Présentation du transcompilateur Qiskit basé sur l'IA

from qiskit import QuantumCircuit
from qiskit.circuit.random import random_circuit
from qiskit.transpiler import generate_preset_pass_manager
from qiskit_ibm_runtime import QiskitRuntimeService, SamplerV2
from qiskit_ibm_transpiler import generate_ai_pass_manager
from qiskit_aer import AerSimulator
from qiskit_aer.noise import NoiseModel, depolarizing_error
import matplotlib.pyplot as plt
from statistics import mean, stdev
import time
import logging

seed = 42


def transpile_with_metrics(pass_manager, circuit):
    """Transpile a circuit and return the result along with key metrics."""
    start = time.time()
    qc_out = pass_manager.run(circuit)
    elapsed = time.time() - start

    depth_2q = qc_out.depth(lambda x: x.operation.num_qubits == 2)
    gate_count = qc_out.size()

    return qc_out, {
        "depth_2q": depth_2q,
        "gate_count": gate_count,
        "time_s": round(elapsed, 3),
    }


def remap_to_contiguous(tqc):
    """Remap a transpiled circuit to use contiguous qubit indices.

    Transpiled circuits target specific physical qubits (e.g., qubit 45, 67)
    on a large backend. This remaps them to 0, 1, 2, ... so Aer only
    simulates the active qubits."""
    active = sorted(
        {tqc.find_bit(q).index for inst in tqc.data for q in inst.qubits}
    )
    qubit_map = {old: new for new, old in enumerate(active)}
    new_qc = QuantumCircuit(len(active))
    for inst in tqc.data:
        old_indices = [tqc.find_bit(q).index for q in inst.qubits]
        new_qc.append(inst.operation, [qubit_map[i] for i in old_indices])
    return new_qc


def build_mirror_circuit(tqc, simulate=True):
    """Build a mirror circuit: U followed by U-dagger, with measurements.

    The expected output is always |0...0>, so measuring the survival
    probability reveals how much noise each transpilation strategy adds.

    Args:
        tqc: A transpiled circuit.
        simulate: If True (default), remap to contiguous qubits so Aer
            only simulates the active qubits. If False, keep the full
            physical layout for hardware execution."""
    if simulate:
        tqc = remap_to_contiguous(tqc)
    mirror = tqc.compose(tqc.inverse())
    mirror.measure_all()
    return mirror


def print_summary(results):
    """Print a summary of each metric as mean +/- stdev across all circuits,
    along with the mean percentage improvement of AI over Default."""
    metrics = [
        ("Depth 2Q", "Depth 2Q (Default)", "Depth 2Q (AI)"),
        ("Gate Count", "Gate Count (Default)", "Gate Count (AI)"),
        ("Time (s)", "Time (Default)", "Time (AI)"),
    ]
    header = (
        f"{'Metric':<12}{'Default (mean +/- std)':>24}"
        f"{'AI (mean +/- std)':>22}{'AI % improvement':>22}"
    )
    print(header)
    print("-" * len(header))
    for label, col_def, col_ai in metrics:
        defaults = [r[col_def] for r in results]
        ais = [r[col_ai] for r in results]
        pct = [(d - a) / d * 100 for d, a in zip(defaults, ais)]
        default_str = f"{mean(defaults):.1f} +/- {stdev(defaults):.1f}"
        ai_str = f"{mean(ais):.1f} +/- {stdev(ais):.1f}"
        pct_str = f"{mean(pct):+.1f}% +/- {stdev(pct):.1f}%"
        print(f"{label:<12}{default_str:>24}{ai_str:>22}{pct_str:>22}")


def plot_metrics_and_pct(results, title_prefix):
    """Plot metric comparisons and percentage improvement of AI over Default."""
    qubits = [r["Qubits"] for r in results]
    metrics = [
        ("Depth 2Q (Default)", "Depth 2Q (AI)", "Two-Qubit Depth"),
        ("Gate Count (Default)", "Gate Count (AI)", "Gate Count"),
        ("Time (Default)", "Time (AI)", "Transpilation Time"),
    ]

    # Row 1: raw metric comparison
    fig, axs = plt.subplots(1, 3, figsize=(21, 5))
    fig.suptitle(
        f"{title_prefix}: Metric Comparison",
        fontsize=15,
        fontweight="bold",
        y=1.02,
    )
    for ax, (col_def, col_ai, label) in zip(axs, metrics):
        ax.plot(qubits, [r[col_def] for r in results], "o-", label="Default")
        ax.plot(qubits, [r[col_ai] for r in results], "s-", label="AI")
        ax.set_title(label)
        ax.set_xlabel("Number of Qubits")
        ax.set_ylabel(label)
        ax.legend()
    plt.tight_layout()
    plt.show()

    # Row 2: percentage improvement
    fig, axs = plt.subplots(1, 3, figsize=(21, 5))
    fig.suptitle(
        f"{title_prefix}: % Improvement of AI over Default",
        fontsize=15,
        fontweight="bold",
        y=1.02,
    )
    for ax, (col_def, col_ai, label) in zip(axs, metrics):
        pct = [(r[col_def] - r[col_ai]) / r[col_def] * 100 for r in results]
        ax.axhline(
            0, color="#1f77b4", linewidth=2, label="Default (baseline)"
        )
        ax.plot(qubits, pct, "s-", color="#ff7f0e", label="AI")
        ax.fill_between(qubits, 0, pct, alpha=0.15, color="#ff7f0e")
        ax.set_title(label)
        ax.set_xlabel("Number of Qubits")
        ax.set_ylabel("% Improvement")
        ax.legend()
    plt.tight_layout()
    plt.show()


# Suppress verbose AI-powered transpiler logs
logging.getLogger(
    "qiskit_ibm_transpiler.wrappers.ai_local_synthesis"
).setLevel(logging.WARNING)

num_circuits_sim = 20
depth_sim = 4
qubit_range_sim = list(range(6, 26))

circuits_sim = [
    # We have only two qubit gates, as those test how well the
    # transpiler can optimize the circuit.
    random_circuit(
        num_qubits=n,
        depth=depth_sim,
        max_operands=2,
        num_operand_distribution={2: 1},
        seed=seed + i,
    )
    for i, n in enumerate(qubit_range_sim)
]

print(
    f"Created {len(circuits_sim)} circuits with qubit counts "
    f"from {qubit_range_sim[0]} to {qubit_range_sim[-1]}"
)
circuits_sim[0].draw(output="mpl", fold=-1)

service = QiskitRuntimeService()
backend = service.least_busy(
    min_num_qubits=100, operational=True, simulator=False
)


pm_default_sim = generate_preset_pass_manager(
    optimization_level=3,
    backend=backend,
    seed_transpiler=seed,
)

results_sim = []

for i, qc in enumerate(circuits_sim):
    n = qubit_range_sim[i]

    qc_default, m_default = transpile_with_metrics(pm_default_sim, qc)

    # Create a fresh AI pass manager each iteration to avoid stale layout state
    pm_ai = generate_ai_pass_manager(
        optimization_level=1,
        ai_optimization_level=3,
        backend=backend,
    )
    qc_ai, m_ai = transpile_with_metrics(pm_ai, qc)

    results_sim.append(
        {
            "Qubits": n,
            "Depth 2Q (Default)": m_default["depth_2q"],
            "Depth 2Q (AI)": m_ai["depth_2q"],
            "Gate Count (Default)": m_default["gate_count"],
            "Gate Count (AI)": m_ai["gate_count"],
            "Time (Default)": m_default["time_s"],
            "Time (AI)": m_ai["time_s"],
        }
    )

print_summary(results_sim)

plot_metrics_and_pct(results_sim, "Small-Scale Random Circuits")

# Use the 10-qubit circuit (index where qubits == 10)
idx_10q = qubit_range_sim.index(10)

qc_10q = circuits_sim[idx_10q]
qc_default_10q, _ = transpile_with_metrics(pm_default_sim, qc_10q)

pm_ai = generate_ai_pass_manager(
    optimization_level=1,
    ai_optimization_level=3,
    backend=backend,
)
qc_ai_10q, _ = transpile_with_metrics(pm_ai, qc_10q)

tqc_methods = {
    "Default": qc_default_10q,
    "AI": qc_ai_10q,
}

print(
    f"Default: depth {qc_default_10q.depth()}, gates {qc_default_10q.size()}"
)
print(f"AI:      depth {qc_ai_10q.depth()}, gates {qc_ai_10q.size()}")

# Build a simple depolarizing noise model
noise_model = NoiseModel()
noise_model.add_all_qubit_quantum_error(
    depolarizing_error(0.001, 1),
    ["sx", "x", "rz"],  # ~0.1% per 1Q gate
)
noise_model.add_all_qubit_quantum_error(
    depolarizing_error(0.01, 2),
    ["cx", "ecr"],  # ~1% per 2Q gate
)

aer_sim = AerSimulator(noise_model=noise_model)

shots = 10000
survival_probs = {}

for method, tqc in tqc_methods.items():
    mirror = build_mirror_circuit(tqc, simulate=True)

    sampler = SamplerV2(mode=aer_sim)
    job = sampler.run([mirror], shots=shots)
    counts = job.result()[0].data.meas.get_counts()

    all_zeros = "0" * mirror.num_qubits
    survival = counts.get(all_zeros, 0) / shots
    survival_probs[method] = survival
    print(
        f"{method:8s}  P(|00...0>) = {survival:.4f}  "
        f"({counts.get(all_zeros, 0)}/{shots})"
    )

# Plot 1 - P(|0...0>), the probability of an erroneous (non-zero) outcome.
# A lower bar means the transpilation introduced less noise.
error_probs = {method: 1 - p for method, p in survival_probs.items()}

fig, ax = plt.subplots(figsize=(6, 4))
ax.bar(
    error_probs.keys(),
    error_probs.values(),
    color=["steelblue", "coral"],
)
ax.set_ylabel("1 - P(|0...0>)")
ax.set_title("Mirror Circuit Error (10-qubit, Aer Simulator)")
ax.set_ylim(0, 1)
plt.tight_layout()
plt.show()

# -------------------------Step 1-------------------------
num_circuits_hw = 25
depth_hw = 8
qubit_range_hw = list(range(26, 51))

circuits_hw = [
    # We have only two qubit gates, as those test how well the
    # transpiler can optimize the circuit.
    random_circuit(
        num_qubits=n,
        depth=depth_hw,
        max_operands=2,
        num_operand_distribution={2: 1},
        seed=seed + i,
    )
    for i, n in enumerate(qubit_range_hw)
]

print(
    f"Created {len(circuits_hw)} circuits with qubit counts "
    f"from {qubit_range_hw[0]} to {qubit_range_hw[-1]}"
)

# -------------------------Step 2-------------------------
pm_default = generate_preset_pass_manager(
    optimization_level=3,
    backend=backend,
    seed_transpiler=seed,
)

results_hw = []

for i, qc in enumerate(circuits_hw):
    n = qubit_range_hw[i]

    qc_default, m_default = transpile_with_metrics(pm_default, qc)

    # Create a fresh AI pass manager each iteration to avoid stale layout state
    pm_ai = generate_ai_pass_manager(
        optimization_level=1,
        ai_optimization_level=3,
        backend=backend,
    )
    qc_ai, m_ai = transpile_with_metrics(pm_ai, qc)

    results_hw.append(
        {
            "Qubits": n,
            "Depth 2Q (Default)": m_default["depth_2q"],
            "Depth 2Q (AI)": m_ai["depth_2q"],
            "Gate Count (Default)": m_default["gate_count"],
            "Gate Count (AI)": m_ai["gate_count"],
            "Time (Default)": m_default["time_s"],
            "Time (AI)": m_ai["time_s"],
        }
    )

print_summary(results_hw)

plot_metrics_and_pct(results_hw, "Large-Scale Random Circuits")

# -------------------------Step 3-------------------------
# Build mirror circuits from the 26-qubit case
idx_26q = qubit_range_hw.index(26)

qc_26q = circuits_hw[idx_26q]
qc_default_26q, _ = transpile_with_metrics(pm_default, qc_26q)

pm_ai = generate_ai_pass_manager(
    optimization_level=1,
    ai_optimization_level=3,
    backend=backend,
)
qc_ai_26q, _ = transpile_with_metrics(pm_ai, qc_26q)

mirror_default_hw = build_mirror_circuit(qc_default_26q, simulate=False)
mirror_ai_hw = build_mirror_circuit(qc_ai_26q, simulate=False)

# Re-transpile to basis gates (the inverse can introduce gates like sxdg)
pm_basis = generate_preset_pass_manager(
    optimization_level=0,
    backend=backend,
)
mirror_default_hw = pm_basis.run(mirror_default_hw)
mirror_ai_hw = pm_basis.run(mirror_ai_hw)

print(
    f"Mirror circuit (Default): depth {mirror_default_hw.depth()}, "
    f"gates {mirror_default_hw.size()}"
)
print(
    f"Mirror circuit (AI):      depth {mirror_ai_hw.depth()}, "
    f"gates {mirror_ai_hw.size()}"
)

# Submit to real hardware
sampler_hw = SamplerV2(mode=backend)
sampler_hw.options.environment.job_tags = ["TUT_AITI"]

shots_hw = 500000
job_hw = sampler_hw.run([mirror_default_hw, mirror_ai_hw], shots=shots_hw)
print(f"Job submitted: {job_hw.job_id()}")

# -------------------------Step 4-------------------------
result_hw = job_hw.result()

survival_probs_hw = {}
for i, method in enumerate(["Default", "AI"]):
    counts = result_hw[i].data.meas.get_counts()
    mirror = [mirror_default_hw, mirror_ai_hw][i]
    all_zeros = "0" * mirror.num_qubits
    survival = counts.get(all_zeros, 0) / shots_hw
    survival_probs_hw[method] = survival
    print(
        f"{method:8s}  P(|00...0>) = {survival:.4f}  "
        f"({counts.get(all_zeros, 0)}/{shots_hw})"
    )

# Plot 1 - P(|0...0>), the probability of an erroneous (non-zero) outcome.
# A lower bar means the transpilation introduced less noise.
error_probs_hw = {method: 1 - p for method, p in survival_probs_hw.items()}

fig, ax = plt.subplots(figsize=(6, 4))
ax.bar(
    error_probs_hw.keys(),
    error_probs_hw.values(),
    color=["steelblue", "coral"],
)
ax.set_ylabel("1 - P(|0...0>)")
ax.set_title(f"Mirror Circuit Error (26-qubit, {backend.name})")
ax.set_ylim(0, 1)
plt.tight_layout()
plt.show()