Open-source tools from my research. See also my GitHub.
ADSALA
Architecture and Data-Structure Aware Linear Algebra
A machine learning library that optimizes the runtime of BLAS routines on modern multi-core systems. ADSALA predicts the optimal number of threads for each BLAS operation based on matrix dimensions and system architecture, achieving 1.5-3.0x speedups over default configurations.
- Platforms: Intel (MKL) and AMD (BLIS) multi-core systems
- Operations: All BLAS Level 3 (GEMM, TRSM, SYMM, etc.)
- Award: Best Poster at IEEE IPDPS 2024
- [Interactive Demo] [Paper (IPDPS’24)] [Paper (IPDPS’23)]
NN-xTB
Neural Network Extended Tight-Binding
A Hamiltonian-preserving scheme that augments semi-empirical quantum chemistry (GFN2-xTB) with neural-network-predicted parameter shifts. Achieves DFT-level accuracy at near semi-empirical cost while preserving physical interpretability, analytic long-range limits, and self-consistency.
- Accuracy: WTMAD-2 of 4 kcal/mol on GMTKN55 (vs. 25 for GFN2-xTB)
- Speed: Near-xTB computational cost
- Benchmarks: GMTKN55, rMD17 force predictions
- [Preprint]