Applicant
Prof. Dr. Fakher F. Assaad
Institut für Theoretische Physik
Universität Würzburg
Project Overview
The Algorithms for Lattice Fermions (ALF) project is an open‑source initiative that provides state‑of‑the‑art implementations of auxiliary‑field quantum Monte Carlo methods for studying correlated electron systems. Used worldwide across the solid‑state physics community, ALF enables researchers to simulate a broad range of quantum models with high numerical precision and robust parallel performance. With this KONWIHR project, we aim to significantly expand ALF’s capabilities, usability, and performance to meet the needs of modern high‑performance computing architectures.
A major pillar of the project is ease of use, centered on the further development of pyALF, the Python interface to ALF. Many collaborators, especially experimentalists, work primarily in Python and Jupyter notebooks. By exposing ALF directly as a Python module, enabling model and observable definitions in Python, shipping binaries with the package, and automating essential accuracy checks, we aim to make ALF accessible to non‑experts. This includes automated detection of numerical instabilities, autocorrelation analysis, warmup estimation, and identification of problematic fat‑tailed distributions. The redesigned interface will be tested and refined through user feedback, including during the ALF School in October 2026.
The second component of the project focuses on GPU acceleration of key numerical kernels, in particular the QR decomposition with pivoting, currently the dominant computational bottleneck in ALF. Recent advances, such as the GPU‑optimized implementation in the RandLAPACK library, now make it possible to replace the traditional LAPACK routine (ZGEQP3) and evaluate performance gains on modern architectures. Another line of attack is to carry out measurements of time-displaced correlation function on the GPU. This is a stand-alone task that follows the leading scaling of the computational cost.
Finally, we plan to implement a Hybrid Monte Carlo (HMC) algorithm with stochastic forces, a method particularly well suited for GPU acceleration. While ALF already includes an exact‑force HMC implementation, the stochastic‑force variant opens the door to efficient sparse‑matrix computations that map naturally onto GPUs. This will allow users, where the physics permits, to achieve substantial speedups simply by selecting the new algorithm at runtime. Implementing this method requires foundational work: reviewing existing ALF components, restructuring the codebase toward a more modular library design, and integrating the new update scheme alongside existing algorithms.
The project is organized in two phases: first, refining pyALF in close collaboration with the user community and testing GPU‑acceleration schemes; second, implementing the HMC algorithm.
Overall, this KONWIHR project strengthens ALF as a sustainable, modular, user‑friendly, and GPU‑ready research software ecosystem, ensuring that the solid‑state physics community can fully exploit emerging high‑performance computing architectures.