Get started

Run the computational workflow in Linux or WSL. All commands below start in the repository root. On WSL, a checkout on the Linux filesystem gives better performance for the many small measurement writes.

Install the environment

Install an MPI Fortran compiler, Make, and BLAS/LAPACK. For example, on Ubuntu or Debian:

sudo apt-get update
sudo apt-get install -y git make gfortran openmpi-bin libopenmpi-dev libblas-dev liblapack-dev

Create the supplied Python environment with Conda:

conda env create -f benchmarks/paper/environment.yml
conda activate bafqmc
python3 scripts/doctor.py

The environment contains NumPy, Matplotlib, and QuSpin for ED. An existing Python 3.11 environment can instead install the two requirements files:

python3 -m pip install -r benchmarks/paper/requirements.txt -r benchmarks/paper/requirements-ed.txt

The Makefiles use mpifort with GNU Fortran by default. An activated Intel MPI/MKL environment is also supported. Build configuration and optional overrides are described in the shared numerical guide. For a separate QuSpin environment, pass --python-ed /path/to/python to the root reproduction command and environment doctor.

Check the installation

make check
python3 reproduce.py --mode smoke

The smoke workflow builds both solvers, executes small Monte Carlo runs, computes exact references, and checks analysis. It writes results under benchmarks/paper/output/smoke/. For another run, select a fresh destination:

python3 reproduce.py --mode smoke --output /tmp/bafqmc-smoke-second

The smoke cases use reduced sampling and ED basis sizes. They verify the installation before production.

Reproduce all benchmark data

python3 reproduce.py --plan
python3 reproduce.py --output benchmarks/paper/output/paper-run

The second command builds and runs BAFQMC, computes the ED references, then creates tables and figures from the new measurements. It includes all 22 points: the 15 main benchmark points and seven supplemental points. Each production point uses 100000 measurement bins and the original input settings. For the reference desktop, budget 12–24 hours, 16 GiB RAM with about 8 GiB available to the calculation, and 8 GiB free disk. See measured resource estimates for details.

Outputs in the selected directory include:

runs/                                raw BAFQMC data and ED results
progress.json                        completed stages and timing
observables.csv                      means, SEM, and ED values
block_means.csv                       blocked measurements
records.json                         results with parameters and provenance
figures/benchmark_combined.pdf        main benchmark
figures/benchmark_attractive.pdf      supplemental benchmark
environment.json                     environment and selected-case metadata

Continue after an interruption with the same output directory and options:

python3 reproduce.py --output benchmarks/paper/output/paper-run --resume

Completed stages are checked and reused. The interrupted stage is rerun from its original inputs. The root runner manages scratch automatically; use --work-dir /path/to/empty/linux-scratch to choose its location.

To compute a selected part of the paper:

python3 reproduce.py --scope main
python3 reproduce.py --scope supplement
python3 reproduce.py --scope main --model pairing

The standalone python3 reproduce.py command always computes the full campaign by default. --mode plot is available to inspect the stored processed data. A list of all stages and data conventions is in the reproduction guide.

Start your own calculation

Tell your agent the Hamiltonian, lattice size, temperature, observables, desired parameter scan, and available resources. Have it read AGENTS.md and follow the custom-campaign recipes. The solver-specific runners accept separate manifests, so research inputs can be developed independently of the published benchmark package.