暗黙的溶媒モデルを用いた電子構造シミュレーション用テンプレートのデプロイと実行
このテンプレートは、Cleveland Clinicとの共同開発によるものであり、暗黙的溶媒中の分子の基底状態エネルギーおよび溶媒和自由エネルギーを計算するワークフローで構成されています [1]。これらのシミュレーションは、サンプルベース量子対角化（SQD）法 [2-6] と、溶媒の積分方程式形式分極連続体モデル（IEF-PCM）[7] に基づいています。
このガイドでは、溶質としてメタノール分子、溶媒として水（連続誘電体媒体として近似）を用いてテンプレートを利 用します。メタノールの電子相関効果を考慮しながら、計算コストと精度のバランスを保つために、SQD IEF-PCMでシミュレートする活性空間には 、、および孤立対軌道のみを含めます。この軌道選択は、C[2s,2p]、O[2s,2p]、H[1s]の原子軌道成分を用いた原子価活性空間（AVAS）法によって行われ、14電子12軌道（14e,12o）の活性空間が得られます。参照軌道は、cc-pvdz基底関数を用いた閉殻Hartree-Fock法で計算されます。
ワークフローの紹介
このインタラクティブガイドでは、このファンクションテンプレートをQiskit Serverlessにアップロードし、サンプルワークロードを実行する方法を示します。テンプレートはQiskitパターンとして4つのステップで構成されています。
1. 入力の収集と問題のマッピング
このステップでは、分子の幾何学的形状、選択された活性空間、溶媒和モデル、LUCJオプション、およびSQDオプションを入力として受け取ります。そして、Hartree-Fock（HF）IEF-PCMデータを含むPySCFチェックポイントファイルを生成します。このデータは、ワークフローのSQD部分で使用されます。ワークフローのLUCJ部分については、入力セクションでPySCF FCIDUMPフォーマットで内部的に保存される気相HFデータも生成されます。
HF気相シミュレーションの情報と活性空間の定義が入力として使用されます。重要なのは、エラー抑制、ショット数、回路トランスパイラの最適化レベル、および量子ビットレイアウトに関する入力セクションからのユーザー定義情報も使用される点です。
定義された活性空間内の1電子積分と2電子積分が生成されます。これらの積分は、LUCJ回路のパラメータ化に使用するt2振幅を返す古典的CCSD計算に用いられます。
2. 回路の最適化
LUCJ回路は、ターゲットハードウェア向けのISA回路にトランスパイルされます。次に、実行を管理するためのデフォルトのエラー軽減オプションセットでSamplerプリミティブがインスタンス化されます。
3. 回路の実行
LUCJ計算は各測定のビット列を返します。これらのビット列は、研究対象システムの電子配置に対応します。ビット列はその後、後処理の入力として使用されます。
4. SQDを用いた後処理
この最終ステップでは、HF IEF-PCM情報を含むPySCFチェックポイントファイル、LUCJによって予測された電子配置を表すビット列、および入力セクションで選択されたユーザー定義SQDオプションを入力として受け取ります。出力として、最低エネルギーバッチのSQD IEF-PCM全エネルギーと対応する溶媒和自由エネルギーが生成されます。
オプション
このテンプレートでは、LUCJ回路を生成するためのオプションとSQD実行パラメータを指定する必要があります。
LUCJオプション
LUCJ量子回路が実行されると、分子系の確率分布から計算基底状態を表すサンプルセットが生成されます。LUCJ回路の深さと表現力のバランスを保つために、反対スピンのスピン軌道に対応する量子ビットには、これらの量子ビットが単一アンシラ量子ビットを介して隣接している場合に2量子ビットゲートが適用されます。このアプローチをヘビーヘックストポロジーを持つIBMハードウェアに実装するために、同じスピンのスピン軌道を表す量子ビットはラインとして接続され、ターゲットハードウェアのヘビーヘックス接続性によりジグザグ形状になります。一方、反対スピンのスピン軌道を表す量子ビットは、4量子ビットごとにのみ接続されます。
必要なオプションの詳細を表示するにはクリックしてください:
ユーザーは、SQD IEF-PCMファンクションの lucj_options セクションで、この_ジグザグ_パターンを満たす量子ビットに対応する initial_layout 配列を提供する必要があります。メタノールのSQD IEF-PCM (14e,12o)/cc-pvdz シミュレーションの場合、Eagle R3 QPUの主対角線に対応する 初期量子ビットレイアウトを選択しました。initial_layout 配列の最初の12要素 [0, 14, 18, 19, 20, 33, 39, 40, 41, 53, 60, 61, ...] はアルファスピン軌道に対応します。最後の12要素 [... 2, 3, 4, 15, 22, 23, 24, 34, 43, 44, 45, 54] はベータスピン軌道に対応します。
重要なのは、ユーザーが number_of_shots を決定する必要があることです。これはLUCJ回路の測定回数に対応します。S-CORE手順の最初のステップが基底状態占有数分布の初期近似を得るために正しい粒子セクター内のサンプルに依存するため、ショット数は十分に大きい必要があります。
ショット数はシステムおよびハードウェアに大きく依存しますが、非共有結合、フラグメントベース、および暗黙的溶媒のSQD研究では、以下のガイドラインに従うことで化学精度に到達できることが示されています。
- 分子軌道数が16未満（スピン軌道32未満）のシステムには 20,000 ～ 200,000ショット
- 分子軌道数が16 ～ 18のシステムには 200,000ショット
- 分子軌道数が18を超えるシステムには 200,000 ～ 2,000,000ショット
必要なショット数は、研究対象システムのスピン軌道数と、研究対象システム内の選択された活性空間に対応するヒルベルト空間のサイズに影響されます。一般に、ヒルベルト空間が小さいインスタンスは必要なショット数が少なくて済みます。その他の利 用可能なLUCJオプションとして、回路トランスパイラの最適化レベルとエラー抑制オプションがあります。これらのオプションも必要なショット数と結果の精度に影響することにご注意ください。
SQDオプション
SQDシミュレーションの重要なオプションには sqd_iterations、number_of_batches、および samples_per_batch があります。一般的に、バッチあたりのサンプル数が少ない場合は、バッチ数（number_of_batches）とS-COREの反復回数（sqd_iterations）を増やすことで補えます。バッチ数が増えると、配置部分空間のより多くのバリエーションをサンプリングできます。最低エネルギーバッチがシステムの基底状態エネルギーの解として採用されるため、バッチ数が多いほど統計的に優れた結果が得られます。S-COREの追加の反復によって、正しい粒子セクターのサンプル数が少ない場合でも、元のLUCJ分布からより多くの配置を回復できます。これにより、バッチあたりのサンプル数を減らすことが可能になります。
SQDオプションの設定に関する詳細情報を表示するにはクリックしてください:
別の戦 略として、バッチあたりのサンプル数を増やすことが考えられます。これにより、正しい粒子空間にある初期LUCJサンプルのほとんどがS-CORE手順で使用されるようになり、個々の部分空間が十分に多様な電子配置を包含します。その結果、必要なS-COREステップ数が減り、バッチあたりのサンプル数が十分に多い場合は2～3回のSQD反復のみで済みます。ただし、バッチあたりのサンプル数が多いと、各対角化ステップの計算コストが高くなります。したがって、sqd_iterations、number_of_batches、および samples_per_batch を最適に選択することで、SQDシミュレーションにおける精度と計算コストのバランスを達成できます。
SQD IEF-PCMの研究では、3回のS-CORE反復を使用した場合、以下のガイドラインに従うことで化学精度に到達できることが示されています。
- メタノールSQD IEF-PCM (14e,12o) シミュレーションではバッチあたり600サンプル
- メチルアミンSQD IEF-PCM (14e,13o) シミュレーションではバッチあたり1500サンプル
- 水SQD IEF-PCM (8e,23o) シミュレーションではバッチあたり6000サンプル
- エタノールSQD IEF-PCM (20e,18o) シミュレーションではバッチあたり16000サンプル
LUCJで必要なショット数と同様に、S-CORE手順で使用されるバッチあたりの必要サンプル数は、システムおよびハードウェアに大きく依存します。上記の例は、必要なバッチあたりのサンプル数のベンチマークの初期点を推定するために使用できます。必要なバッチあたりのサンプル数の系統的なベン チマークに関するチュートリアルはこちらをご覧ください。
テンプレートSQD IEF-PCMファンクションのデプロイと実行
# Added by doQumentation — required packages for this notebook
!pip install -q ffsim numpy pyscf qiskit qiskit-addon-sqd qiskit-ibm-catalog qiskit-ibm-runtime qiskit-serverless solve-solvent
認証
qiskit-ibm-catalog を使用して、APIキー（トークン）で QiskitServerless に認証してください。トークンはIBM Quantum Platformのダッシュボードで確認できます。これにより、選択したファンクションをアップロードまたは実行するためのサーバーレスクライアントをインスタンス化できます。
from qiskit_ibm_catalog import QiskitServerless
serverless = QiskitServerless(
channel="ibm_quantum_platform",
instance="INSTANCE_CRN",
token="YOUR_API_KEY" # Use the 44-character API_KEY you created and saved from the IBM Quantum Platform Home dashboard
)
オプションとして、save_account() を使用してローカル環境に認証情報を保存できます（IBM Cloudアカウントの設定ガイドを参照）。これにより認証情報は QiskitRuntimeService.save_account() と同じファイルに書き込まれることに注意してください。
QiskitServerless.save_account(token="YOUR_API_KEY", channel="ibm_quantum_platform", instance="INSTANCE_CRN")
アカウントが保存済みの場合、認証にトークンを提供する必要はありません。
from qiskit_ibm_catalog import QiskitServerless
serverless = QiskitServerless()
テンプレートのアップロード
カスタムQiskit Functionをアップロードするには、まずファンクションのソースコードを定義する QiskitFunction オブジェクトをインスタンス化する必要があります。タイトルを指定することで、リモートクラスター上でファンクションを識別できます。メインエントリーポイントは if __name__ == "__main__" を含むファイルです。ワークフローに追加のソースファイルが必要な場合は、エントリーポイントと一緒にアップロードされる作業ディレクトリを定義できます。
from qiskit_ibm_catalog import QiskitFunction
template = QiskitFunction(
title="sqd_pcm_template",
entrypoint="sqd_pcm_entrypoint.py",
working_dir="./source_files/", # all files in this directory will be uploaded
dependencies=[
"ffsim==0.0.54",
"pyscf==2.9.0",
"qiskit_addon_sqd==0.10.0",
],
)
print(template)
QiskitFunction(sqd_pcm_template)
インスタンスの準備が整ったら、サーバーレスにアップロードします。
serverless.upload(template)
QiskitFunction(sqd_pcm_template)
プログラムが正常にアップロードされたことを確認するには、serverless.list() を使用してください。
serverless.list()
[QiskitFunction(sqd_pcm_template),
QiskitFunction(hamiltonian_simulation_template)]
テンプレートのリモートロードと実行
ファンクションテンプレートがアップロードされたので、Qiskit Serverlessを使用してリモートで実行できます。まず、名前でテンプレートをロードします。
template = serverless.load("sqd_pcm_template")
print(template)
QiskitFunction(sqd_pcm_template)
次に、SQD-IEF PCMのドメインレベルの入力でテンプレートを実行します。この例では、メタノールベースのワークロードを指定しています。
molecule = {
"atom": """
O -0.04559 -0.75076 -0.00000;
C -0.04844 0.65398 -0.00000;
H 0.85330 -1.05128 -0.00000;
H -1.08779 0.98076 -0.00000;
H 0.44171 1.06337 0.88811;
H 0.44171 1.06337 -0.88811
""", # Must be specified
"basis": "cc-pvdz", # default is "sto-3g"
"spin": 0, # default is 0
"charge": 0, # default is 0
"verbosity": 0, # default is 0
"number_of_active_orb": 12, # Must be specified
"number_of_active_alpha_elec": 7, # Must be specified
"number_of_active_beta_elec": 7, # Must be specified
"avas_selection": [
"%d O %s" % (k, x) for k in [0] for x in ["2s", "2px", "2py", "2pz"]
]
+ ["%d C %s" % (k, x) for k in [1] for x in ["2s", "2px", "2py", "2pz"]]
+ ["%d H 1s" % k for k in [2, 3, 4, 5]], # default is None
}
solvent_options = {
"method": "IEF-PCM", # other available methods are COSMO, C-PCM, SS(V)PE, see https://manual.q-chem.com/5.4/topic_pcm-em.html
"eps": 78.3553, # value for water
}
lucj_options = {
"initial_layout": [
0,
14,
18,
19,
20,
33,
39,
40,
41,
53,
60,
61,
2,
3,
4,
15,
22,
23,
24,
34,
43,
44,
45,
54,
],
"dynamical_decoupling_choice": True,
"twirling_choice": True,
"number_of_shots": 200000,
"optimization_level": 2,
}
sqd_options = {
"sqd_iterations": 3,
"number_of_batches": 10,
"samples_per_batch": 1000,
"max_davidson_cycles": 200,
}
backend_name = "ibm_sherbrooke"
job = template.run(
backend_name=backend_name,
molecule=molecule,
solvent_options=solvent_options,
lucj_options=lucj_options,
sqd_options=sqd_options,
)
print(job.job_id)
39f8fb70-79b2-43ca-b723-84e6b6135821
ジョブの詳細なステータスを確認します。
import time
t0 = time.time()
status = job.status()
if status == "QUEUED":
print(f"time = {time.time()-t0:.2f}, status = QUEUED")
while True:
status = job.status()
if status == "QUEUED":
continue
print(f"time = {time.time()-t0:.2f}, status = {status}")
if status == "DONE" or status == "ERROR":
break
time = 2.35, status = DONE
ジョブの実行中に、logger.info の出力から作成されたログを取得できます。これらのログは、SQD IEF-PCMワークフローの進行状況に関する有益な情報を提供します。例えば、同じスピン軌道の接続情報や、ハードウェアで実行するための最終ISA回路の2量子ビット深さなどの情報が得られます。
print(job.logs())
ジョブの結果を呼び出すと、結果が利用可能になるまでプログラムの残りの部分がブロックされます。ジョブが完了したら、結果を取得できます。結果には、溶媒和自由エネルギー、最低エネルギーバッチに関する情報、最低エネルギー値、およびソルバーの総実行時間などの有益な情報が含まれます。
result = job.result()
result
{'total_energy_hist': array([[-115.14768518, -115.1368396 , -114.19181692, -115.13745429,
-115.1445012 , -114.19673326, -115.1547003 , -114.20563866,
-115.13748344, -115.14764974],
[-115.15768392, -115.15850126, -115.15857275, -115.15770916,
-115.15801684, -115.15822125, -115.15833521, -115.15844051,
-115.15735538, -115.15862354],
[-115.15795148, -115.15847925, -115.15856677, -115.15811156,
-115.15815602, -115.15785171, -115.1583672 , -115.1585533 ,
-115.15833528, -115.15808791]]),
'spin_squared_value_hist': array([[5.37327508e-03, 1.32981759e-02, 1.36214922e-02, 8.84413615e-03,
7.26723578e-03, 1.94875195e-02, 3.03153152e-03, 6.07543106e-03,
1.04951849e-02, 5.36529204e-03],
[6.39397528e-04, 1.36814350e-04, 9.09054260e-05, 5.99361358e-04,
3.64261739e-04, 2.54905866e-04, 2.32540370e-04, 1.53181990e-04,
7.23519739e-04, 6.80737671e-05],
[4.53776416e-04, 1.63043449e-04, 1.05317263e-04, 3.82912836e-04,
3.41047803e-04, 5.18620393e-04, 2.06819142e-04, 1.17086537e-04,
2.32357159e-04, 4.26071537e-04]]),
'solvation_free_energy_hist': array([[-0.00725018, -0.00743955, -0.01132905, -0.0073377 , -0.00722221,
-0.01136705, -0.00719279, -0.01072829, -0.00733404, -0.00725961],
[-0.00719252, -0.00718315, -0.00718074, -0.00719325, -0.00717703,
-0.00718391, -0.00718354, -0.00717928, -0.00719887, -0.0071801 ],
[-0.00719351, -0.00718255, -0.00718198, -0.00718429, -0.00718349,
-0.00718329, -0.0071882 , -0.00718363, -0.00718549, -0.00718814]]),
'occupancy_hist': [[array([0.99712298, 0.99278936, 0.99083163, 0.97328469, 0.98959809,
0.98922134, 0.720333 , 0.25683194, 0.01939338, 0.02840332,
0.00946988, 0.0327204 ]),
array([0.99712298, 0.99278936, 0.99083163, 0.97328469, 0.98959809,
0.98922134, 0.720333 , 0.25683194, 0.01939338, 0.02840332,
0.00946988, 0.0327204 ])],
[array([0.9959042 , 0.9922607 , 0.99018862, 0.99265843, 0.98927447,
0.9900833 , 0.99403876, 0.00989025, 0.01120814, 0.01137717,
0.01152871, 0.01158725]),
array([0.9959042 , 0.9922607 , 0.99018862, 0.99265843, 0.98927447,
0.9900833 , 0.99403876, 0.00989025, 0.01120814, 0.01137717,
0.01152871, 0.01158725])],
[array([0.99590079, 0.99222193, 0.99016753, 0.99265045, 0.98927264,
0.99007179, 0.99407207, 0.00986684, 0.01125181, 0.01141439,
0.01150733, 0.01160243]),
array([0.99590079, 0.99222193, 0.99016753, 0.99265045, 0.98927264,
0.99007179, 0.99407207, 0.00986684, 0.01125181, 0.01141439,
0.01150733, 0.01160243])]],
'lowest_energy_batch': 2,
'lowest_energy_value': -115.1585667736213,
'solvation_free_energy': -0.007181981952470838,
'sci_solver_total_duration': 493.997501373291,
'metadata': {'resources_usage': {'RUNNING: MAPPING': {'CPU_TIME': 6.080063343048096},
'RUNNING: OPTIMIZING_FOR_HARDWARE': {'CPU_TIME': 1.999896764755249},
'RUNNING: WAITING_FOR_QPU': {'CPU_TIME': 6.2850868701934814},
'RUNNING: EXECUTING_QPU': {'QPU_TIME': 21.639373540878296},
'RUNNING: POST_PROCESSING': {'CPU_TIME': 495.40831995010376}},
'num_iterations_executed': 3}}
結果のメタデータには、各ワークロードに必要なQPUおよびCPU時間をより正確に見積もるためのリソース使用量サマリーが含まれていることに注意してください（この例はダミーデバイスで実行されているため、実際のリソース使用時間は異なる場合があります）。ジョブが完了すると、ログ出力全体が利用可能になります。
print(job.logs())
2025-06-27 08:42:41,358 INFO job_manager.py:531 -- Runtime env is setting up.
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:45,015: Starting runtime service
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:45,621: Backend: ibm_sherbrooke
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:46,809: Initializing molecule object
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,599: Performing CCSD
Parsing /tmp/ray/session_2025-06-27_08-42-13_898146_1/runtime_resources/working_dir_files/_ray_pkg_4bc93dcc58c04b91/output_sqd_pcm/2025-06-27_08-42-45.fcidump.txt
Overwritten attributes get_ovlp get_hcore of <class 'pyscf.scf.hf_symm.SymAdaptedRHF'>
/usr/local/lib/python3.11/site-packages/pyscf/gto/mole.py:1293: UserWarning: Function mol.dumps drops attribute energy_nuc because it is not JSON-serializable
warnings.warn(msg)
/usr/local/lib/python3.11/site-packages/pyscf/gto/mole.py:1293: UserWarning: Function mol.dumps drops attribute intor_symmetric because it is not JSON-serializable
warnings.warn(msg)
converged SCF energy = -115.049680672847
E(CCSD) = -115.1519910037652 E_corr = -0.1023103309180226
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,694: Same spin orbital connections: [(0, 1), (1, 2), (2, 3), (3, 4), (4, 5), (5, 6), (6, 7), (7, 8), (8, 9), (9, 10), (10, 11)]
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:51,694: Opposite spin orbital connections: [(0, 0), (4, 4), (8, 8)]
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,718: Optimization level: 2, ops: OrderedDict([('rz', 2438), ('sx', 1496), ('ecr', 766), ('x', 185), ('measure', 24), ('barrier', 1)]), depth: 391
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,736: Two-qubit gate depth: 94
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:53,737: Submitting sampler job
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:54,273: Job ID: d1f5j3lqbivc73ebqpj0
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:42:54,313: Job Status: QUEUED
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,813: Starting configuration recovery iteration 0
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,841: Batch 0 subspace dimension: 531441
2025-06-27 08:43:24,844 INFO worker.py:1588 -- Using address 172.17.16.124:6379 set in the environment variable RAY_ADDRESS
2025-06-27 08:43:24,847 INFO worker.py:1723 -- Connecting to existing Ray cluster at address: 172.17.16.124:6379...
2025-06-27 08:43:24,876 INFO worker.py:1908 -- Connected to Ray cluster. View the dashboard at [1m[32mhttp://172.17.16.124:8265 [39m[22m
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,945: Batch 1 subspace dimension: 519841
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,950: Batch 2 subspace dimension: 543169
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,955: Batch 3 subspace dimension: 532900
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,960: Batch 4 subspace dimension: 534361
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,964: Batch 5 subspace dimension: 531441
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,969: Batch 6 subspace dimension: 540225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,974: Batch 7 subspace dimension: 524176
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,979: Batch 8 subspace dimension: 537289
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:43:24,983: Batch 9 subspace dimension: 540225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,006: Lowest energy batch: 6
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Lowest energy value: -115.15470029849135
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Corresponding g_solv value: -0.0071927910374866375
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:09,007: Starting configuration recovery iteration 1
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,564: Batch 0 subspace dimension: 413449
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,572: Batch 1 subspace dimension: 399424
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,578: Batch 2 subspace dimension: 438244
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,583: Batch 3 subspace dimension: 422500
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,589: Batch 4 subspace dimension: 409600
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,596: Batch 5 subspace dimension: 404496
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,601: Batch 6 subspace dimension: 410881
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,605: Batch 7 subspace dimension: 442225
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,611: Batch 8 subspace dimension: 409600
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:48:40,618: Batch 9 subspace dimension: 405769
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Lowest energy batch: 9
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Lowest energy value: -115.15862353596414
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,917: Corresponding g_solv value: -0.0071800982859467006
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,918: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:49:54,918: Starting configuration recovery iteration 2
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,501: Batch 0 subspace dimension: 399424
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,508: Batch 1 subspace dimension: 412164
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,514: Batch 2 subspace dimension: 432964
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,519: Batch 3 subspace dimension: 400689
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,524: Batch 4 subspace dimension: 432964
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,529: Batch 5 subspace dimension: 418609
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,533: Batch 6 subspace dimension: 418609
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,538: Batch 7 subspace dimension: 425104
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,543: Batch 8 subspace dimension: 404496
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:50:25,548: Batch 9 subspace dimension: 429025
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,900: Lowest energy batch: 2
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,900: Lowest energy value: -115.1585667736213
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: Corresponding g_solv value: -0.007181981952470838
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: -----------------------------------
sqd_pcm_entrypoint.run_function:INFO:2025-06-27 08:51:37,901: SCI_solver totally takes: 493.997501373291 seconds
次のステップ
推奨事項
- ハミルトニアンシミュレーションのファンクションテンプレート構築ガイドを参照してください
- このテンプレートのソースファイルは GitHub でご確認いただけます
参考文献
[1] Danil Kaliakin, Akhil Shajan, Fangchun Liang, and Kenneth M. Merz Jr. Implicit Solvent Sample-Based Quantum Diagonalization, The Journal of Physical Chemistry B, 2025, DOI: 10.1021/acs.jpcb.5c01030
[2] Javier Robledo-Moreno, et al., Chemistry Beyond Exact Solutions on a Quantum-Centric Supercomputer, arXiv:2405.05068 [quant-ph].
[3] Jeffery Yu, et al., Quantum-Centric Algorithm for Sample-Based Krylov Diagonalization, arXiv:2501.09702 [quant-ph].
[4] Keita Kanno, et al., Quantum-Selected Configuration Interaction: classical diagonalization of Hamiltonians in subspaces selected by quantum computers, arXiv:2302.11320 [quant-ph].
[5] Kenji Sugisaki, et al., Hamiltonian simulation-based quantum-selected configuration interaction for large-scale electronic structure calculations with a quantum computer, arXiv:2412.07218 [quant-ph].
[6] Mathias Mikkelsen, Yuya O. Nakagawa, Quantum-selected configuration interaction with time-evolved state, arXiv:2412.13839 [quant-ph].
[7] Herbert, John M. Dielectric continuum methods for quantum chemistry. WIREs Computational Molecular Science, 2021, 11, 1759-0876.
[8] Saki, A. A.; Barison, S.; Fuller, B.; Garrison, J. R.; Glick, J. R.; Johnson, C.; Mezzacapo, A.; Robledo-Moreno, J.; Rossmannek, M.; Schweigert, P. et al. Qiskit addon: sample-based quantum diagonalization, 2024; https://github.com/Qiskit/qiskit-addon-sqd
[9] Asun, Q.; Zhang, X.; Banerjee, S.; Bao, P.; Barbry, M.; Blunt, N. S.; Bogdanov, N. A.; Booth, G. H.; Chen, J.; Cui, Z.-H. PySCF: Python-based Simulations of Chemistry Framework, 2025; https://github.com/pyscf/pyscf
[10] Kevin J. Sung; et al., FFSIM: Faster simulations of fermionic quantum circuits, 2024. https://github.com/qiskit-community/ffsim
%%writefile ./source_files/__init__.py
%%writefile ./source_files/solve_solvent.py
# This code is part of a Qiskit project.
#
# (C) Copyright IBM and Cleveland Clinic 2025
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""Functions for the study of fermionic systems."""
from __future__ import annotations
import warnings
import numpy as np
# DSK Add imports needed for CASCI wrapper
from pyscf import ao2mo, scf, fci
from pyscf.mcscf import avas, casci
from pyscf.solvent import pcm
from pyscf.lib import chkfile, logger
from qiskit_addon_sqd.fermion import (
SCIState,
bitstring_matrix_to_ci_strs,
_check_ci_strs,
)
# DSK Below is the modified CASCI kernel compatible with SQD.
# It utilizes the "fci.selected_ci.kernel_fixed_space"
# as well as enables passing the "batch" and "max_davidson"
# input arguments from "solve_solvent".
# The "batch" contains the CI addresses corresponding to subspaces
# derived from LUCJ and S-CORE calculations.
# The "max_davidson" controls the maximum number of cycles of Davidson's algorithm.
# pylint: disable = unused-argument
def kernel(casci_object, mo_coeff=None, ci0=None, verbose=logger.NOTE, envs=None):
"""CASCI solver compatible with SQD.
Args:
casci_object: CASCI or CASSCF object.
In case of SQD, only CASCI instance is currently incorporated.
mo_coeff : ndarray
orbitals to construct active space Hamiltonian.
In context of SQD, these are either AVAS mo_coeff
or all of the MOs (with option to exclude core MOs).
ci0 : ndarray or custom types FCI solver initial guess.
For SQD the usage of ci0 was not tested.
For external FCI-like solvers, it can be
overloaded different data type. For example, in the state-average
FCI solver, ci0 is a list of ndarray. In other solvers such as
DMRGCI solver, SHCI solver, ci0 are custom types.
kwargs:
envs: dict
In case of SQD this option was not explored,
but in principle this can facilitate the incorporation of the external solvers.
The variable envs is created (for PR 807) to passes MCSCF runtime
environment variables to SHCI solver. For solvers which do not
need this parameter, a kwargs should be created in kernel method
and "envs" pop in kernel function.
"""
if mo_coeff is None:
mo_coeff = casci_object.mo_coeff
if ci0 is None:
ci0 = casci_object.ci
log = logger.new_logger(casci_object, verbose)
t0 = (logger.process_clock(), logger.perf_counter())
log.debug("Start CASCI")
ncas = casci_object.ncas
nelecas = casci_object.nelecas
# The start of SQD version of kernel
# DSK add the read of configurations for batch
ci_strs_sqd = casci_object.batch
# DSK add the input for the maximum number of cycles of Davidson's algorithm
max_davidson = casci_object.max_davidson
# DSK add electron up and down count and norb = ncas
n_up = nelecas[0]
n_dn = nelecas[1]
norb = ncas
# DSK Eigenstate solver info
sqd_verbose = verbose
# DSK ERI read
eri_cas = ao2mo.restore(1, casci_object.get_h2eff(), casci_object.ncas)
t1 = log.timer("integral transformation to CAS space", *t0)
# DSK 1e integrals
h1eff, energy_core = casci_object.get_h1eff()
log.debug("core energy = %.15g", energy_core)
t1 = log.timer("effective h1e in CAS space", *t1)
if h1eff.shape[0] != ncas:
raise RuntimeError(
"Active space size error. nmo=%d ncore=%d ncas=%d" # pylint: disable=consider-using-f-string
% (mo_coeff.shape[1], casci_object.ncore, ncas)
)
# DSK fcisolver needs to be defined in accordance with SQD
# in this software stack it is done in the "solve_solvent" portion of the code.
myci = casci_object.fcisolver
e_cas, sqdvec = fci.selected_ci.kernel_fixed_space(
myci,
h1eff,
eri_cas,
norb,
(n_up, n_dn),
ci_strs=ci_strs_sqd,
verbose=sqd_verbose,
max_cycle=max_davidson,
)
# DSK fcivec is the general name for CI vector assigned by PySCF.
# Depending on type of solver it is either FCI or SCI vector.
# In case of sqd we can call it "sqdvec" for clarity.
# Nonetheless, for further processing PySCF expects
# this data structure to be called fcivec, regardless of the used solver.
fcivec = sqdvec
t1 = log.timer("CI solver", *t1)
e_tot = energy_core + e_cas
# Returns either standard CASCI data or SQD data. Return depends on "sqd_run" True/False.
return e_tot, e_cas, fcivec
# Replace standard CASCI kernel with the SQD-compatible CASCI kernel defined above
casci.kernel = kernel
def solve_solvent(
bitstring_matrix: tuple[np.ndarray, np.ndarray] | np.ndarray,
/,
myeps: float,
mysolvmethod: str,
myavas: list,
num_orbitals: int,
*,
spin_sq: int | None = None,
max_davidson: int = 100,
verbose: int | None = 0,
checkpoint_file: str,
) -> tuple[float, SCIState, list[np.ndarray], float]:
"""Approximate the ground state given molecular integrals and a set of electronic configurations.
Args:
bitstring_matrix: A set of configurations defining the subspace onto which the Hamiltonian
will be projected and diagonalized. This is a 2D array of ``bool`` representations of bit
values such that each row represents a single bitstring. The spin-up configurations
should be specified by column indices in range ``(N, N/2]``, and the spin-down
configurations should be specified by column indices in range ``(N/2, 0]``, where ``N``
is the number of qubits.
(DEPRECATED) The configurations may also be specified by a length-2 tuple of sorted 1D
arrays containing unsigned integer representations of the determinants. The two lists
should represent the spin-up and spin-down orbitals, respectively.
To build PCM model PySCF needs the structure of the molecule. Hence, the electron integrals
(hcore and eri) are not enough to form IEF-PCM simulation. Instead the "start.chk" file is used.
This workflow also requires additional information about solute and solvent,
which is reflected by additional arguments below
myeps: Dielectric parameter of the solvent.
mysolvmethod: Solvent model, which can be IEF-PCM, COSMO, C-PCM, SS(V)PE,
see https://manual.q-chem.com/5.4/topic_pcm-em.html
At the moment only IEF-PCM was tested.
In principle two other models from PySCF "solvent" module can be used as well,
namely SMD and polarizable embedding (PE).
The SMD and PE were not tested yet and their usage requires addition of more
input arguments for "solve_solvent".
myavas: This argument allows user to select active space in solute with AVAS.
The corresponding list should include target atomic orbitals.
If myavas=None, then active space selected based on number of orbitals
derived from ci_strs.
It is assumed that if myavas=None, then the target calculation is either
a) corresponds to full basis case.
b) close to full basis case and only few core orbitals are excluded.
num_orbitals: Number of orbitals, which is essential when myavas = None.
In AVAS case number of orbitals and electrons is derived by AVAS procedure itself.
spin_sq: Target value for the total spin squared for the ground state.
If ``None``, no spin will be imposed.
max_davidson: The maximum number of cycles of Davidson's algorithm
verbose: A verbosity level between 0 and 10
checkpoint_file: Name of the checkpoint file
NOTE: For now open shell functionality is not supported in SQD PCM calculations.
Hence, at the moment solve_solvent does not include open_shell as one of the arguments.
Returns:
- Minimum energy from SCI calculation
- The SCI ground state
- Average occupancy of the alpha and beta orbitals, respectively
- Expectation value of spin-squared
- Solvation free energy
"""
# Unlike the "solve_fermion", the "solve_solvent" utilizes the "checkpoint" file to
# get the starting HF information, which means that "solve_solvent" does not accept
# "hcore" and "eri" as the input arguments.
# Instead "hcore" and "eri" are generated inside of the custom SQD-compatible
# CASCI kernel (defined above).
# The generation of "hcore" and "eri" is based on the information from "checkpoint" file
# as well as "myavas" and "num_orbitals" input arguments.
# DSK this part handles addresses and is identical to "solve_fermion"
if isinstance(bitstring_matrix, tuple):
warnings.warn(
"Passing the input determinants as integers is deprecated. "
"Users should instead pass a bitstring matrix defining the subspace.",
DeprecationWarning,
stacklevel=2,
)
ci_strs = bitstring_matrix
else:
# This will become the default code path after the deprecation period.
ci_strs = bitstring_matrix_to_ci_strs(bitstring_matrix, open_shell=False)
ci_strs = _check_ci_strs(ci_strs)
num_up = format(ci_strs[0][0], "b").count("1")
num_dn = format(ci_strs[1][0], "b").count("1")
# DSK assign verbosity
verbose_ci = verbose
# DSK add information about solute and solvent.
# Since PCM model needs the information about the structure of the molecule
# one cannot use only FCIDUMP. Instead converged HF data can be passed from "checkpoint" file
# along with "mol" object containing the geometry and other information about the solute.
############################################
# This section is specific to "solve_solvent" and is not present in "solve_fermion".
# In case of "solve_fermion" the "eri" and "hcore" are passed directly to
# "fci.selected_ci.kernel_fixed_space".
# In case of "solve_solvent" the incorporation of the polarizable continuum model
# requires utilization of "CASCI.with_solvent"
# data object from PySCF, where underlying CASCI.base_kernel has to be replaced
# with SQD-compatible version.
# Due to these differences in the implementation the "solve_solvent" recovers
# the converged mean field results and "molecule" object from "checkpoint" file
# (instead of using FCIDUMP),
# followed by passing of solute, solvent, and active space information to "CASCI.with_solvent".
# This includes the initiation of "mol", "cm", "mf", and "mc" data structures.
mol = chkfile.load_mol(checkpoint_file)
# DSK Initiation of the solvent model
cm = pcm.PCM(mol)
cm.eps = myeps # solute eps value
cm.method = mysolvmethod # IEF-PCM, COSMO, C-PCM, SS(V)PE,
# see https://manual.q-chem.com/5.4/topic_pcm-em.html
# DSK Read-in converged RHF solution
scf_result_dic = chkfile.load(checkpoint_file, "scf")
mf = scf.RHF(mol).PCM(cm)
mf.__dict__.update(scf_result_dic)
# Identify the active space based on the user input of AVAS or number of orbitals and electrons
if myavas is not None:
orbs = myavas
avas_obj = avas.AVAS(mf, orbs, with_iao=True)
avas_obj.kernel()
ncas, nelecas, _, _, _ = (
avas_obj.ncas,
avas_obj.nelecas,
avas_obj.mo_coeff,
avas_obj.occ_weights,
avas_obj.vir_weights,
)
else:
ncas = num_orbitals
nelecas = (num_up, num_dn)
# Initiate the "CASCI.with_solvent" object
mc = casci.CASCI(mf, ncas=ncas, nelecas=nelecas).PCM(cm)
# Replace mo_coeff with ones produced by AVAS if AVAS is utilized
if myavas is not None:
mc.mo_coeff = avas_obj.mo_coeff
# Read-in the configuration interaction subspace derived from LUCJ and S-CORE
mc.batch = ci_strs
# Assign number of maximum Davidson steps
mc.max_davidson = max_davidson
####### The definition of "fcisolver" object is identical to "solve_fermion" case ########
myci = fci.selected_ci.SelectedCI()
if spin_sq is not None:
myci = fci.addons.fix_spin_(myci, ss=spin_sq)
mc.fcisolver = myci
mc.verbose = verbose_ci
#########################################################################################
# Initiate the "CASCI.with_solvent" simulation with SQD-compatible based CASCI kernel.
mc_result = mc.kernel()
# Get data out of the "CASCI.with_solvent" object
e_sci = mc_result[0]
sci_vec = mc_result[2]
# Here we get additional output comparing to "solve_fermion",
# which is the solvation free energy (G_solv)
g_solv = mc.with_solvent.e
#####################################################
# The remainder of the code in solve_solvent is nearly identical to solve_fermion code.
# However, there are two exceptions in "solve_solvent":
# 1) The dm2 is currently not computed, but can be included if needed
# 2) e_sci is directly output as the result of CASCI.with_solvent object.
# Hence, the two following lines of code are not present in "solve_solvent"
# comparing to the "solve_fermion" code:
# dm2 = myci.make_rdm2(sci_vec, norb, (num_up, num_dn))
# e_sci = np.einsum("pr,pr->", dm1, hcore) + 0.5 * np.einsum("prqs,prqs->", dm2, eri)
# Calculate the avg occupancy of each orbital
dm1 = myci.make_rdm1s(sci_vec, ncas, (num_up, num_dn))
avg_occupancy = [np.diagonal(dm1[0]), np.diagonal(dm1[1])]
# Compute total spin
spin_squared = myci.spin_square(sci_vec, ncas, (num_up, num_dn))[0]
# Convert the PySCF SCIVector to internal format. We access a private field here,
# so we assert that we expect the SCIVector output from kernel_fixed_space to
# have its _strs field populated with alpha and beta strings.
assert isinstance(sci_vec._strs[0], np.ndarray) and isinstance(sci_vec._strs[1], np.ndarray)
assert sci_vec.shape == (len(sci_vec._strs[0]), len(sci_vec._strs[1]))
sci_state = SCIState(
amplitudes=np.array(sci_vec),
ci_strs_a=sci_vec._strs[0],
ci_strs_b=sci_vec._strs[1],
)
return e_sci, sci_state, avg_occupancy, spin_squared, g_solv
%%writefile ./source_files/sqc_pcm_entrypoint.py
# This code is part of a Qiskit project.
#
# (C) Copyright IBM and Cleveland Clinic 2025
#
# This code is licensed under the Apache License, Version 2.0. You may
# obtain a copy of this license in the LICENSE.txt file in the root directory
# of this source tree or at http://www.apache.org/licenses/LICENSE-2.0.
#
# Any modifications or derivative works of this code must retain this
# copyright notice, and modified files need to carry a notice indicating
# that they have been altered from the originals.
"""
SQD-PCM Function Template source code.
"""
from pathlib import Path
from typing import Any
from datetime import datetime
import os
import sys
import json
import logging
import time
import traceback
import numpy as np
import ffsim
from pyscf import gto, scf, mcscf, ao2mo, tools, cc
from pyscf.lib import chkfile
from pyscf.mcscf import avas
from pyscf.solvent import pcm
from qiskit import QuantumCircuit, QuantumRegister
from qiskit.transpiler.preset_passmanagers import generate_preset_pass_manager
from qiskit.primitives import BackendSamplerV2
from qiskit_addon_sqd.counts import counts_to_arrays
from qiskit_addon_sqd.configuration_recovery import recover_configurations
from qiskit_addon_sqd.fermion import bitstring_matrix_to_ci_strs
from qiskit_addon_sqd.subsampling import postselect_and_subsample
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2
from qiskit_serverless import get_arguments, save_result, distribute_task, get, update_status, Job
current_dir = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, current_dir)
from solve_solvent import solve_solvent # pylint: disable=wrong-import-position
logger = logging.getLogger(__name__)
def run_function(
backend_name: str,
molecule: dict,
solvent_options: dict,
sqd_options: dict,
lucj_options: dict | None = None,
**kwargs,
) -> dict[str, Any]:
"""
Main entry point for the SQD-PCM (Polarizable Continuum Model) workflow.
This function encapsulates the end-to-end execution of the algorithm.
Args:
backend_name: Identifier for the target backend, required for all
workflows that access IBM Quantum hardware.
molecule: dictionary with molecule information:
- "atom" (str): required field, follows pyscf specification for atomic geometry.
For example, for methanol the value would be::
'''
O -0.04559 -0.75076 -0.00000;
C -0.04844 0.65398 -0.00000;
H 0.85330 -1.05128 -0.00000;
H -1.08779 0.98076 -0.00000;
H 0.44171 1.06337 0.88811;
H 0.44171 1.06337 -0.88811;
'''
- "number_of_active_orb" (int): required field
- "number_of_active_alpha_elec" (int): required field
- "number_of_active_beta_elec" (int): required field
- "basis" (str): optional field, default is "sto-3g"
- "verbosity" (int): optional field, default is 0
- "charge" (int): optional field, default is 0
- "spin" (int): optional field, default is 0
- "avas_selection" (list[str] | None): optional field, default is None
solvent_options: dictionary with solvent options information:
- "method" (str): required field. Method for computing solvent reaction field
for the PCM. Accepted values are: "IEF-PCM", "COSMO",
"C-PCM", "SS(V)PE", see https://manual.q-chem.com/5.4/topic_pcm-em.html
- "eps" (float): required field. Dielectric constant of the solvent in the PCM.
sqd_options: dictionary with sqd options information:
- "sqd_iterations" (int): required field.
- "number_of_batches" (int): required field.
- "samples_per_batch" (int): required field.
- "max_davidson_cycles" (int): required field.
lucj_options: optional dictionary with lucj options information:
- "optimization_level" (int): optional field, default is 2
- "initial_layout" (list[int]): optional field, default is None
- "dynamical_decoupling" (bool): optional field, default is True
- "twirling" (bool): optional field, default is True
- "number_of_shots" (int): optional field, default is 10000
**kwargs
Optional keyword arguments to customize behavior. Existing kwargs include:
- "files_name" (str): optional name for output files (enabled for local testing)
- "testing_backend" (FakeBackendV2): optional fake backend instance to bypass
qiskit runtime service instantiation (enabled for local testing)
- "count_dict_file_name" (str): path to a count dict file to bypass primitive
execution and jump directly to SQD section (enabled for local testing)
Returns:
The function should return the execution results as a dictionary with string keys.
This is to ensure compatibility with ``qiskit_serverless.save_result``.
"""
# Preparation Step: Input validation.
# Do this at the top of the function definition so it fails early if any required
# arguments are missing or invalid.
# Molecule parsing
# Required:
geo = molecule["atom"]
num_active_orb = molecule["number_of_active_orb"]
num_active_alpha = molecule["number_of_active_alpha_elec"]
num_active_beta = molecule["number_of_active_beta_elec"]
# Optional:
input_basis = molecule.get("basis", "sto-3g")
input_verbosity = molecule.get("verbosity", 0)
input_charge = molecule.get("charge", 0)
input_spin = molecule.get("spin", 0)
myavas = molecule.get("avas_selection", None)
# Solvent options parsing
myeps = solvent_options["eps"]
mymethod = solvent_options["method"]
# LUCJ options parsing
if lucj_options is None:
lucj_options = {}
opt_level = lucj_options.get("optimization_level", 2)
initial_layout = lucj_options.get("initial_layout", None)
use_dd = lucj_options.get("dynamical_decoupling", True)
use_twirling = lucj_options.get("twirling", True)
num_shots = lucj_options.get("number_of_shots", True)
# SQD options parsing
iterations = sqd_options["sqd_iterations"]
n_batches = sqd_options["number_of_batches"]
samples_per_batch = sqd_options["samples_per_batch"]
max_davidson_cycles = sqd_options["max_davidson_cycles"]
# kwarg parsing (local testing)
testing_backend = kwargs.get("testing_backend", None)
count_dict_file_name = kwargs.get("count_dict_file_name", None)
files_name = kwargs.get("files_name", datetime.now().strftime("%Y-%m-%d_%H-%M-%S"))
output_path = Path.cwd() / "output_sqd_pcm"
output_path.mkdir(exist_ok=True)
datafiles_name = str(output_path) + "/" + files_name
# --
# Preparation Step: Qiskit Runtime & primitive configuration for
# execution on IBM Quantum hardware.
if testing_backend is None:
# Initialize Qiskit Runtime Service
logger.info("Starting runtime service")
service = QiskitRuntimeService(
channel=os.environ["QISKIT_IBM_CHANNEL"],
instance=os.environ["IBM_CLOUD_INSTANCE"],
token=os.environ["your-API_KEY"], # Use the 44-character API_KEY you created and saved from the IBM Quantum Platform Home dashboard
)
backend = service.backend(backend_name)
logger.info(f"Backend: {backend.name}")
# Set up sampler and corresponding options
sampler = SamplerV2(backend)
sampler.options.dynamical_decoupling.enable = use_dd
sampler.options.twirling.enable_measure = False
sampler.options.twirling.enable_gates = use_twirling
sampler.options.default_shots = num_shots
else:
backend = testing_backend
logger.info(f"Testing backend: {backend.name}")
# Set up backend sampler.
# This doesn't allow running with twirling and dd
sampler = BackendSamplerV2(backend=testing_backend)
# Once the preparation steps are completed, the algorithm can be structured following a
# Qiskit Pattern workflow:
# https://docs.quantum.ibm.com/guides/intro-to-patterns
# --
# Step 1: Map
# In this step, input arguments are used to construct relevant quantum circuits and operators
start_mapping = time.time()
update_status(Job.MAPPING)
# Initialize the molecule object (pyscf)
logger.info("Initializing molecule object")
mol = gto.Mole()
mol.build(
atom=geo,
basis=input_basis,
verbose=input_verbosity,
charge=input_charge,
spin=input_spin,
symmetry=False,
) # Not tested for symmetry calculations
cm = pcm.PCM(mol)
cm.eps = myeps
cm.method = mymethod
mf = scf.RHF(mol).PCM(cm)
# Generation of checkpoint file for the solute and solvent
# which will be used reused in all subsequent sections
checkpoint_file_name = str(datafiles_name + ".chk")
mf.chkfile = checkpoint_file_name
mf.kernel()
# Read-in the information about the molecule
mol = chkfile.load_mol(checkpoint_file_name)
# Read-in RHF data
scf_result_dic = chkfile.load(checkpoint_file_name, "scf")
mf = scf.RHF(mol)
mf.__dict__.update(scf_result_dic)
# LUCJ uses isolated solute
mf.kernel()
# Initialize orbital selection based on user input
if myavas is not None:
orbs = myavas
avas_out = avas.AVAS(mf, orbs, with_iao=True)
avas_out.kernel()
ncas, nelecas = (avas_out.ncas, avas_out.nelecas)
else:
ncas = num_active_orb
nelecas = (
num_active_alpha,
num_active_beta,
)
# LUCJ Step:
# Generate active space
mc = mcscf.CASCI(mf, ncas=ncas, nelecas=nelecas)
if myavas is not None:
mc.mo_coeff = avas_out.mo_coeff
mc.batch = None
# Reliable and most convenient way to do the CCSD on only the active space
# is to create the FCIDUMP file and then run the CCSD calculation only on the
# orbitals stored in the FCIDUMP file.
h1e_cas, ecore = mc.get_h1eff()
h2e_cas = ao2mo.restore(1, mc.get_h2eff(), mc.ncas)
fcidump_file_name = str(datafiles_name + ".fcidump.txt")
tools.fcidump.from_integrals(
fcidump_file_name,
h1e_cas,
h2e_cas,
ncas,
nelecas,
nuc=ecore,
ms=0,
orbsym=[1] * ncas,
)
logger.info("Performing CCSD")
# Read FCIDUMP and perform CCSD on only active space
mf_as = tools.fcidump.to_scf(fcidump_file_name)
mf_as.kernel()
mc_cc = cc.CCSD(mf_as)
mc_cc.kernel()
mc_cc.t1 # pylint: disable=pointless-statement
t2 = mc_cc.t2
n_reps = 2
norb = ncas
if myavas is not None:
nelec = (int(nelecas / 2), int(nelecas / 2))
else:
nelec = nelecas
alpha_alpha_indices = [(p, p + 1) for p in range(norb - 1)]
alpha_beta_indices = [(p, p) for p in range(0, norb, 4)]
logger.info(f"Same spin orbital connections: {alpha_alpha_indices}")
logger.info(f"Opposite spin orbital connections: {alpha_beta_indices}")
# Construct LUCJ op
ucj_op = ffsim.UCJOpSpinBalanced.from_t_amplitudes(
t2, n_reps=n_reps, interaction_pairs=(alpha_alpha_indices, alpha_beta_indices)
)
# Construct circuit
qubits = QuantumRegister(2 * norb, name="q")
circuit = QuantumCircuit(qubits)
circuit.append(ffsim.qiskit.PrepareHartreeFockJW(norb, nelec), qubits)
circuit.append(ffsim.qiskit.UCJOpSpinBalancedJW(ucj_op), qubits)
circuit.measure_all()
end_mapping = time.time()
# --
# Step 2: Optimize
# Transpile circuits to match ISA
start_optimizing = time.time()
update_status(Job.OPTIMIZING_HARDWARE)
pass_manager = generate_preset_pass_manager(
optimization_level=opt_level,
backend=backend,
initial_layout=initial_layout,
)
pass_manager.pre_init = ffsim.qiskit.PRE_INIT
transpiled = pass_manager.run(circuit)
end_optimizing = time.time()
logger.info(
f"Optimization level: {opt_level}, ops: {transpiled.count_ops()}, depth: {transpiled.depth()}"
)
two_q_depth = transpiled.depth(lambda x: x.operation.num_qubits == 2)
logger.info(f"Two-qubit gate depth: {two_q_depth}")
# --
# Step 3: Execute on Hardware
# Submit the underlying Sampler job. Note that this is not the
# actual function job.
if count_dict_file_name is None:
# Submit the LUCJ job
logger.info("Submitting sampler job")
job = sampler.run([transpiled])
logger.info(f"Job ID: {job.job_id()}")
logger.info(f"Job Status: {job.status()}")
start_waiting_qpu = time.time()
while job.status() == "QUEUED":
update_status(Job.WAITING_QPU)
time.sleep(5)
end_waiting_qpu = time.time()
update_status(Job.EXECUTING_QPU)
# Wait until job is complete
result = job.result()
end_executing_qpu = time.time()
pub_result = result[0]
counts_dict = pub_result.data.meas.get_counts()
waiting_qpu_time = end_waiting_qpu - start_waiting_qpu
executing_qpu_time = end_executing_qpu - end_waiting_qpu
else:
# read LUCJ samples from count_dict
logger.info("Skipping sampler, loading counts dict from file")
with open(count_dict_file_name, "r") as file:
count_dict_string = file.read().replace("\n", "")
counts_dict = json.loads(count_dict_string.replace("'", '"'))
waiting_qpu_time = 0
executing_qpu_time = 0
# --
# Step 4: Post-process
start_pp = time.time()
update_status(Job.POST_PROCESSING)
# SQD-PCM section
start = time.time()
# Orbitals, electron, and spin initialization
num_orbitals = ncas
if myavas is not None:
num_elec_a = num_elec_b = int(nelecas / 2)
else:
num_elec_a, num_elec_b = nelecas
spin_sq = input_spin
# Convert counts into bitstring and probability arrays
bitstring_matrix_full, probs_arr_full = counts_to_arrays(counts_dict)
# We set qiskit_serverless to explicitly reserve 1 cpu per thread, as
# the task is CPU-bound and might degrade in performance when sharing
# a core at scale (this might not be the case with smaller examples)
@distribute_task(target={"cpu": 1})
def solve_solvent_parallel(
batches,
myeps,
mysolvmethod,
myavas,
num_orbitals,
spin_sq,
max_davidson,
checkpoint_file,
):
return solve_solvent( # sqd for pyscf
batches,
myeps,
mysolvmethod,
myavas,
num_orbitals,
spin_sq=spin_sq,
max_davidson=max_davidson,
checkpoint_file=checkpoint_file,
)
e_hist = np.zeros((iterations, n_batches)) # energy history
s_hist = np.zeros((iterations, n_batches)) # spin history
g_solv_hist = np.zeros((iterations, n_batches)) # g_solv history
occupancy_hist = []
avg_occupancy = None
num_ran_iter = 0
for i in range(iterations):
logger.info(f"Starting configuration recovery iteration {i}")
# On the first iteration, we have no orbital occupancy information from the
# solver, so we begin with the full set of noisy configurations.
if avg_occupancy is None:
bs_mat_tmp = bitstring_matrix_full
probs_arr_tmp = probs_arr_full
# If we have average orbital occupancy information, we use it to refine the full
# set of noisy configurations
else:
bs_mat_tmp, probs_arr_tmp = recover_configurations(
bitstring_matrix_full, probs_arr_full, avg_occupancy, num_elec_a, num_elec_b
)
# Create batches of subsamples. We post-select here to remove configurations
# with incorrect hamming weight during iteration 0, since no config recovery was performed.
batches = postselect_and_subsample(
bs_mat_tmp,
probs_arr_tmp,
hamming_right=num_elec_a,
hamming_left=num_elec_b,
samples_per_batch=samples_per_batch,
num_batches=n_batches,
)
# Run eigenstate solvers in a loop. This loop should be parallelized for larger problems.
e_tmp = np.zeros(n_batches)
s_tmp = np.zeros(n_batches)
g_solvs_tmp = np.zeros(n_batches)
occs_tmp = []
coeffs = []
res1 = []
for j in range(n_batches):
strs_a, strs_b = bitstring_matrix_to_ci_strs(batches[j])
logger.info(f"Batch {j} subspace dimension: {len(strs_a) * len(strs_b)}")
res1.append(
solve_solvent_parallel(
batches[j],
myeps,
mymethod,
myavas,
num_orbitals,
spin_sq=spin_sq,
max_davidson=max_davidson_cycles,
checkpoint_file=checkpoint_file_name,
)
)
res = get(res1)
for j in range(n_batches):
energy_sci, coeffs_sci, avg_occs, spin, g_solv = res[j]
e_tmp[j] = energy_sci
s_tmp[j] = spin
g_solvs_tmp[j] = g_solv
occs_tmp.append(avg_occs)
coeffs.append(coeffs_sci)
# Combine batch results
avg_occupancy = tuple(np.mean(occs_tmp, axis=0))
# Track optimization history
e_hist[i, :] = e_tmp
s_hist[i, :] = s_tmp
g_solv_hist[i, :] = g_solvs_tmp
occupancy_hist.append(avg_occupancy)
lowest_e_batch_index = np.argmin(e_hist[i, :])
logger.info(f"Lowest energy batch: {lowest_e_batch_index}")
logger.info(f"Lowest energy value: {np.min(e_hist[i, :])}")
logger.info(f"Corresponding g_solv value: {g_solv_hist[i, lowest_e_batch_index]}")
logger.info("-----------------------------------")
num_ran_iter += 1
end_pp = time.time()
end = time.time()
duration = end - start
logger.info(f"SCI_solver totally takes: {duration} seconds")
metadata = {
"resources_usage": {
"RUNNING: MAPPING": {
"CPU_TIME": end_mapping - start_mapping,
},
"RUNNING: OPTIMIZING_FOR_HARDWARE": {
"CPU_TIME": end_optimizing - start_optimizing,
},
"RUNNING: WAITING_FOR_QPU": {
"CPU_TIME": waiting_qpu_time,
},
"RUNNING: EXECUTING_QPU": {
"QPU_TIME": executing_qpu_time,
},
"RUNNING: POST_PROCESSING": {
"CPU_TIME": end_pp - start_pp,
},
},
"num_iterations_executed": num_ran_iter,
}
output = {
"total_energy_hist": e_hist,
"spin_squared_value_hist": s_hist,
"solvation_free_energy_hist": g_solv_hist,
"occupancy_hist": occupancy_hist,
"lowest_energy_batch": lowest_e_batch_index,
"lowest_energy_value": np.min(e_hist[i, :]),
"solvation_free_energy": g_solv_hist[i, lowest_e_batch_index],
"sci_solver_total_duration": duration,
"metadata": metadata,
}
return output
def set_up_logger(my_logger: logging.Logger, level: int = logging.INFO) -> None:
"""Logger setup to communicate logs through serverless."""
log_fmt = "%(module)s.%(funcName)s:%(levelname)s:%(asctime)s: %(message)s"
formatter = logging.Formatter(log_fmt)
# Set propagate to `False` since handlers are to be attached.
my_logger.propagate = False
stream_handler = logging.StreamHandler()
stream_handler.setFormatter(formatter)
my_logger.addHandler(stream_handler)
my_logger.setLevel(level)
# This is the section where `run_function` is called, it's boilerplate code and can be used
# without customization.
if __name__ == "__main__":
# Use serverless helper function to extract input arguments,
input_args = get_arguments()
# Allow to configure logging level
logging_level = input_args.get("logging_level", logging.INFO)
set_up_logger(logger, logging_level)
try:
func_result = run_function(**input_args)
# Use serverless function to save the results that
# will be returned in the job.
save_result(func_result)
except Exception:
save_result(traceback.format_exc())
raise
sys.exit(0)
# This cell is hidden from users. It verifies both source listings are identical then deletes the working folder we created
import shutil
with open("./source_files/sqd_pcm_entrypoint.py") as f1:
with open("./source_files/sqd_pcm_entrypoint.py") as f2:
assert f1.read() == f2.read()
with open("./source_files/solve_solvent.py") as f1:
with open("./source_files/solve_solvent.py") as f2:
assert f1.read() == f2.read()
with open("./source_files/__init__.py") as f1:
with open("./source_files/__init__.py") as f2:
assert f1.read() == f2.read()
shutil.rmtree("./source_files/")