Quantum Monte Carlo Library Achieves Highly Accurate Electronic Structure Calculations

Quantum Monte Carlo methods provide exceptionally accurate simulations of materials, but their computational demands often limit the size and complexity of systems scientists can study. To address this challenge, Emiel Slootman from the University of Twente, Vijay Gopal Chilkuri and Aurelien Delval from Université Paris-Saclay, along with colleagues, developed QMCkl, a new kernel library designed to accelerate these calculations. This library delivers a collection of efficient, portable code modules that form the foundation of Quantum Monte Carlo simulations, covering essential components from atomic orbitals to complex Jastrow factors. By separating the underlying algorithms from the specific hardware used, QMCkl ensures consistent and reproducible results across different computer architectures, while also achieving significant speed improvements in calculating energy and its derivatives, ultimately enabling more complex and accurate materials modelling.

Accurate calculations rely on carefully selected basis sets and pseudopotentials, which simplify computational demands while maintaining accuracy. The development and implementation of these methods often involve specialized software and tools. TREXIO serves as a library for managing data in parallel quantum chemistry calculations, while packages like Gaussian provide comprehensive computational capabilities.

Robust package managers such as Spack, specifically designed for high-performance computing environments, are essential for managing software dependencies. Improving accuracy often involves techniques like backflow transformations and analyzing electronic structure using atomic basin functions and Bader charge analysis. Recognizing the complexity of QMC methods, the team engineered a modular and portable collection of high-performance kernels that implement the core building blocks of these calculations. The library supports a C-compatible Application Programming Interface and the TREXIO standard for input, encompassing essential kernels such as atomic and molecular orbitals, cusp corrections, and the Jastrow factor, alongside necessary derivatives for variational and structural optimization. A key innovation lies in QMCkl’s separation of algorithmic development from hardware-specific tuning.

Scientists write kernels in a clear, pedagogical Fortran version prioritizing correctness and readability, while high-performance computing specialists implement computationally intensive kernels in optimized C. Both versions produce identical numerical results, accessed through the same C API, ensuring consistent numerical behavior across different QMC codes like CHAMP and QMC=Chem. This modular structure promotes clarity, maintainability, and reproducibility, allowing all QMC codes to benefit from verified and optimized routines. Inspired by classical high-performance numerical libraries like BLAS and LAPACK, QMCkl fosters a principle of separation of concerns.

This approach allows researchers to retain full control over their scientific workflows while leveraging optimized performance, and facilitates broad adoption through interoperability with languages including Python, Fortran, and C++. Beyond QMC, the team demonstrated QMCkl’s applicability to deterministic quantum chemistry programs like Quantum Package, enabling reuse of kernels for evaluating Jastrow factors and computing density grids. Key kernels within QMCkl include evaluations of atomic and molecular orbitals, Jastrow correlation factors, and cusp corrections, all crucial for performing variational and structural optimization. A distinctive feature of QMCkl is its separation of concerns, mirroring the design of established numerical libraries like BLAS and LAPACK.

Scientists write kernels initially in Fortran, prioritizing clarity and correctness, while high-performance computing specialists then implement optimized C versions for speed. This modular design extends beyond QMC, enabling deterministic quantum chemistry programs to reuse kernels for evaluating Jastrow factors or computing density grids. By decoupling algorithmic development from low-level performance optimization, QMCkl allows scientists to retain full control over their workflows while leveraging high-performance computing resources. The C-compatible API ensures interoperability with languages such as Python, Fortran, and C++, facilitating broad adoption and enabling adaptation to new hardware without requiring complete code rewrites. This library delivers substantial speedups in evaluating energy and its derivatives, crucial for accurate electronic structure calculations, and facilitates interoperability between different QMC codes and deterministic quantum chemistry frameworks. By separating algorithmic clarity from performance optimisation, QMCkl ensures consistent and reproducible results across various platforms.

The library’s impact extends beyond QMC itself, accelerating tasks such as orbital analysis and visualisation by over 100times compared to standard tools. This increased efficiency makes it suitable for real-time visualisation, large-scale screening, and integration into interactive workflows where speed is paramount. While the library demonstrably improves performance across all stages of a QMC workflow, the authors acknowledge that further optimisation and expansion of the kernel collection remain ongoing areas of development. Future work will likely focus on extending the library’s capabilities and adapting it to emerging computational architectures, thereby broadening the applicability of QMC methods within the wider field of quantum chemistry.