Giuseppe Maria Junior Barca
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Giuseppe Maria Junior Barca has a total of 115 co-authors. Here are the top co-authors by number of shared publications:
- Elise Palethorpe
Australian National University
8 shared publications - Ryan Stocks
Australian National University
7 shared publications - Mark S. Gordon
Department of Chemistry and Ames National Laboratory
6 shared publications - Peter M. W. Gill
The University of Sydney
6 shared publications Alistair P. Rendell
Flinders University, Adelaide, SA 5001, Australia
5 shared publicationsAndrew T. B. Gilbert
Research School of Chemistry , Australian National University , Canberra , Australian Capital Territory 2601 , Australia
4 shared publications
Latest Publications
- Open Access
- 19 August 2026
- Published by American Chemical Society (ACS) in ChemRxiv
High-Performance, High-Angular-Momentum K Engine on Graphics Processing Units
Exchange matrix construction remains one of the main bottlenecks in GPU-accelerated Hartree-Fock and hybrid density-functional calculations, particularly when high-angular-momentum basis functions are used. Traditional recursive electron repulsion integral (ERI) algorithms minimize operation counts on CPUs, but on GPUs they generate a large number of intermediates, leading to high register pressure, memory traffic, and poor hardware utilization. Here, we present a novel GPU engine for the exchange matrix ( K ) construction designed to circumvent these limitations specifically for high-angular-momentum ERIs. The central component is a reformulation of the McMurchie-Davidson (MD) scheme as a fused double-Matrix Multiply (MatMul), allowing the most expensive high-angular-momentum contractions to be mapped onto the high-throughput Tensor Core hardware of modern accelerators while eliminating intermediate global memory traffic. Using low-level control over matrix multiply-accumulate instructions, we further introduce a specialized primitive-screening strategy that reduces unnecessary work while preserving the dense arithmetic structure required for high throughput. By combining this MD-MatMul engine with a Head-Gordon-Pople scheme for low-angular-momentum classes, our hybrid approach provides a high-performance approach for K-matrix formation up to g -functions. Benchmarks on NVIDIA H200 GPUs demonstrate that our K -engine, integrated with our previously developed J -engine, consistently outperforms existing GPU implementations in QUICK and GPU4PySCF. Across a variety of linear and globular systems, we observe speedups ranging from 1.2-3.6x over GPU4PySCF and 10.3-68.7x over QUICK across triple- and quadruple-ζ basis sets.
- 19 August 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessHigh-Performance, High-Angular-Momentum K Engine on Graphics Processing Units
Exchange matrix construction remains one of the main bottlenecks in GPU-accelerated Hartree-Fock and hybrid density-functional calculations, particularly when high-angular-momentum basis functions are used. Traditional recursive electron repulsion integral (ERI) algorithms minimize operation counts on CPUs, but on GPUs they generate a large number of intermediates, leading to high register pressure, memory traffic, and poor hardware utilization. Here, we present a novel GPU engine for the exchange matrix ( K ) construction designed to circumvent these limitations specifically for high-angular-momentum ERIs. The central component is a reformulation of the McMurchie-Davidson (MD) scheme as a fused double-Matrix Multiply (MatMul), allowing the most expensive high-angular-momentum contractions to be mapped onto the high-throughput Tensor Core hardware of modern accelerators while eliminating intermediate global memory traffic. Using low-level control over matrix multiply-accumulate instructions, we further introduce a specialized primitive-screening strategy that reduces unnecessary work while preserving the dense arithmetic structure required for high throughput. By combining this MD-MatMul engine with a Head-Gordon-Pople scheme for low-angular-momentum classes, our hybrid approach provides a high-performance approach for K-matrix formation up to g -functions. Benchmarks on NVIDIA H200 GPUs demonstrate that our K -engine, integrated with our previously developed J -engine, consistently outperforms existing GPU implementations in QUICK and GPU4PySCF. Across a variety of linear and globular systems, we observe speedups ranging from 1.2-3.6x over GPU4PySCF and 10.3-68.7x over QUICK across triple- and quadruple-ζ basis sets.
- Open Access
- 7 August 2026
- Published by Springer Nature in Research Square
NN-Dx: Chemically adaptive dispersion physics for density-functional theory
London dispersion is essential to molecular structure, recognition and condensed-phase stability, yet remains inaccurately modelled by many density-functional approximations. Existing density-free corrections recover the required long-range interaction efficiently, but their accuracy is constrained by fixed response models and globally parametrized damping functions. Here we introduce NN-Dx, a chemically adaptive dispersion correction that embeds equivariant machine learning within a D4-inspired analytical framework. The model predicts environment-dependent modifications to atomic dynamic-polarizability spectra, charge response, Becke-Johnson damping, pairwise scaling and Axilrod–Teller–Muto three-body interactions, while preserving the explicit asymptotic form of the dispersion energy. Across nine density functionals, NN-Dx reduces the mean absolute errors of default DFT-D4 by 70–91% on held-out noncovalent-interaction benchmarks, with seven functionals achieving errors below 0.08 kcal mol −1 . On neutral and charged DES15K subsets, error reductions reach 82–92%. For PBE0 on a common hydrogen, carbon, nitrogen and oxygen subset, NN-Dx achieves a mean absolute error of 0.051 kcal mol −1 , approximately thirteenfold lower than published machine-learned exchange-hole dipole moment models. NN-Dx also reproduces reference binding curves without spurious stationary points and provides analytical derivatives for stable, energy conserving molecular dynamics.
- 7 August 2026
- Published by Springer Nature in Research Square
Open AccessNN-Dx: Chemically adaptive dispersion physics for density-functional theory
London dispersion is essential to molecular structure, recognition and condensed-phase stability, yet remains inaccurately modelled by many density-functional approximations. Existing density-free corrections recover the required long-range interaction efficiently, but their accuracy is constrained by fixed response models and globally parametrized damping functions. Here we introduce NN-Dx, a chemically adaptive dispersion correction that embeds equivariant machine learning within a D4-inspired analytical framework. The model predicts environment-dependent modifications to atomic dynamic-polarizability spectra, charge response, Becke-Johnson damping, pairwise scaling and Axilrod–Teller–Muto three-body interactions, while preserving the explicit asymptotic form of the dispersion energy. Across nine density functionals, NN-Dx reduces the mean absolute errors of default DFT-D4 by 70–91% on held-out noncovalent-interaction benchmarks, with seven functionals achieving errors below 0.08 kcal mol −1 . On neutral and charged DES15K subsets, error reductions reach 82–92%. For PBE0 on a common hydrogen, carbon, nitrogen and oxygen subset, NN-Dx achieves a mean absolute error of 0.051 kcal mol −1 , approximately thirteenfold lower than published machine-learned exchange-hole dipole moment models. NN-Dx also reproduces reference binding curves without spurious stationary points and provides analytical derivatives for stable, energy conserving molecular dynamics.
- Open Access
- 28 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-11
Accelerating Electrostatically-Embedded Fragmentation Methods using Graphics Processing Units
- 28 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-11
Open AccessAccelerating Electrostatically-Embedded Fragmentation Methods using Graphics Processing Units
- Open Access
- 16 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
NN-Dx: Chemically adaptive dispersion physics for density-functional theory
London dispersion remains an error source in routine density-functional theory because common density-functional approximations lack the nonlocal response required for correlated charge fluctuations. We introduce NN-Dx, a density-free neural correction that learns chemically adaptive parameters inside a D4-like model rather than adding a black-box energy. Starting from element-specific isolated-atom dynamic polarizability spectra, an E(3)-equivariant network predicts environmental spectral corrections, adaptive charge-scaling parameters, and local corrections to Becke–Johnson damping, pair scaling and Axilrod–Teller–Muto three-body terms. Trained end-to-end on dispersion residuals, NN-Dx reduces default D4 errors by 66–90% across nine density functionals on the original D4 benchmark, giving sub-0.08 kcal mol -1 mean absolute errors (MAEs) for eight. On charged and neutral DES15K subsets, errors decrease by 85–93%. For PBE0 on a H/C/N/O subset of DES15K, NN-Dx reaches 0.037 kcal mol -1 MAE, about eighteenfold below ML-XDM, while preserving reference stationary-point structure in scanned binding curves.
- 16 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessNN-Dx: Chemically adaptive dispersion physics for density-functional theory
London dispersion remains an error source in routine density-functional theory because common density-functional approximations lack the nonlocal response required for correlated charge fluctuations. We introduce NN-Dx, a density-free neural correction that learns chemically adaptive parameters inside a D4-like model rather than adding a black-box energy. Starting from element-specific isolated-atom dynamic polarizability spectra, an E(3)-equivariant network predicts environmental spectral corrections, adaptive charge-scaling parameters, and local corrections to Becke–Johnson damping, pair scaling and Axilrod–Teller–Muto three-body terms. Trained end-to-end on dispersion residuals, NN-Dx reduces default D4 errors by 66–90% across nine density functionals on the original D4 benchmark, giving sub-0.08 kcal mol -1 mean absolute errors (MAEs) for eight. On charged and neutral DES15K subsets, errors decrease by 85–93%. For PBE0 on a H/C/N/O subset of DES15K, NN-Dx reaches 0.037 kcal mol -1 MAE, about eighteenfold below ML-XDM, while preserving reference stationary-point structure in scanned binding curves.
- Open Access
- 15 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
High-Performance, High-Angular-Momentum K Engine on Graphics Processing Units
Exchange matrix construction remains one of the main bottlenecks in GPU-accelerated Hartree-Fock and hybrid density-functional calculations, particularly when high-angular-momentum basis functions are used. Traditional recursive electron repulsion integral (ERI) algorithms minimize operation counts on CPUs, but on GPUs they generate a large number of intermediates, leading to high register pressure, memory traffic, and poor hardware utilization. Here, we present a novel GPU engine for the exchange matrix ( K ) construction designed to circumvent these limitations specifically for high-angular-momentum ERIs. The central component is a reformulation of the McMurchie-Davidson (MD) scheme as a fused double-Matrix Multiply (MatMul), allowing the most expensive high-angular-momentum contractions to be mapped onto the high-throughput Tensor Core hardware of modern accelerators while eliminating intermediate global memory traffic. Using low-level control over matrix multiply-accumulate instructions, we further introduce a specialized primitive-screening strategy that reduces unnecessary work while preserving the dense arithmetic structure required for high throughput. By combining this MD-MatMul engine with a Head-Gordon-Pople scheme for low-angular-momentum classes, our hybrid approach provides a high-performance approach for K-matrix formation up to g -functions. Benchmarks on NVIDIA H200 GPUs demonstrate that our K -engine, integrated with our previously developed J -engine, consistently outperforms existing GPU implementations in QUICK and GPU4PySCF. Across a variety of linear and globular systems, we observe speedups ranging from 1.1-3.1x over GPU4PySCF and 10.7-58.5x over QUICK across triple- and quadruple-ζ basis sets.
- 15 June 2026
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessHigh-Performance, High-Angular-Momentum K Engine on Graphics Processing Units
Exchange matrix construction remains one of the main bottlenecks in GPU-accelerated Hartree-Fock and hybrid density-functional calculations, particularly when high-angular-momentum basis functions are used. Traditional recursive electron repulsion integral (ERI) algorithms minimize operation counts on CPUs, but on GPUs they generate a large number of intermediates, leading to high register pressure, memory traffic, and poor hardware utilization. Here, we present a novel GPU engine for the exchange matrix ( K ) construction designed to circumvent these limitations specifically for high-angular-momentum ERIs. The central component is a reformulation of the McMurchie-Davidson (MD) scheme as a fused double-Matrix Multiply (MatMul), allowing the most expensive high-angular-momentum contractions to be mapped onto the high-throughput Tensor Core hardware of modern accelerators while eliminating intermediate global memory traffic. Using low-level control over matrix multiply-accumulate instructions, we further introduce a specialized primitive-screening strategy that reduces unnecessary work while preserving the dense arithmetic structure required for high throughput. By combining this MD-MatMul engine with a Head-Gordon-Pople scheme for low-angular-momentum classes, our hybrid approach provides a high-performance approach for K-matrix formation up to g -functions. Benchmarks on NVIDIA H200 GPUs demonstrate that our K -engine, integrated with our previously developed J -engine, consistently outperforms existing GPU implementations in QUICK and GPU4PySCF. Across a variety of linear and globular systems, we observe speedups ranging from 1.1-3.1x over GPU4PySCF and 10.7-58.5x over QUICK across triple- and quadruple-ζ basis sets.
- Open Access
- 20 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
NN-xTB: density functional accuracy at semi empirical speed with neural network extended tight binding
Accurate prediction of molecular structure, energetics, and reactivity requires quantum chemistry methods that remain too costly for large-scale and high-throughput modeling. Semi-empirical tight-binding approaches are much faster, but their largely element-fixed parameters under-resolve local chemical-environment effects, limiting accuracy. We introduce a neural-network extension of extended-tight binding that learns bounded, environment-dependent shifts to Hamiltonian parameters, increasing chemical adaptivity while preserving self-consistency, charge and spin treatment, and correct long-range behavior. We show that, across benchmarks spanning main-group thermochemistry, noncovalent interactions, molecular forces, vibrational spectra and supramolecular complexes, this approach achieves density functional theory accuracy at near tight-binding cost. On the GMTKN55, the weighted mean absolute error decreases from 25.0 to 3.78 kcal/mol relative to GFN2-xTB; molecular force errors are lowest on eight of ten rMD17 molecules and improve by 10-40% relative to leading machine-learned interatomic potentials; and vibrational frequency errors decrease from 200.6 to 12.7 cm-1, with runtime increasing by less than 20%.
- 20 May 2026
- Published by Springer Nature in Nature Communications
- Vol. 17 (1)
Open AccessNN-xTB: density functional accuracy at semi empirical speed with neural network extended tight binding
Accurate prediction of molecular structure, energetics, and reactivity requires quantum chemistry methods that remain too costly for large-scale and high-throughput modeling. Semi-empirical tight-binding approaches are much faster, but their largely element-fixed parameters under-resolve local chemical-environment effects, limiting accuracy. We introduce a neural-network extension of extended-tight binding that learns bounded, environment-dependent shifts to Hamiltonian parameters, increasing chemical adaptivity while preserving self-consistency, charge and spin treatment, and correct long-range behavior. We show that, across benchmarks spanning main-group thermochemistry, noncovalent interactions, molecular forces, vibrational spectra and supramolecular complexes, this approach achieves density functional theory accuracy at near tight-binding cost. On the GMTKN55, the weighted mean absolute error decreases from 25.0 to 3.78 kcal/mol relative to GFN2-xTB; molecular force errors are lowest on eight of ten rMD17 molecules and improve by 10-40% relative to leading machine-learned interatomic potentials; and vibrational frequency errors decrease from 200.6 to 12.7 cm-1, with runtime increasing by less than 20%.
- 4 May 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (10) , 5060-5072
Double-Hybrid, but Not Double-Cost: GPU-Accelerated DHDFT for the COMPAS-3 Data Set of Polybenzenoid Hydrocarbons
We present the first implementation of double hybrid density functional theory (DHDFT) with all major computational steps accelerated on GPUs. To efficiently utilize GPU hardware, we employ the resolution-of-identity (RI) approximation, transforming electron repulsion integral (ERI) computations into large dense matrix multiplications and enabling basis functions up to g angular momentum. We demonstrate revDSD-PBEP86-D4(noFC)/def2-QZVPP calculations on the entire COMPAS-3x data set of ∼39,000 peri-condensed polybenzenoid hydrocarbon isomer geometries (up to 68 atoms) using just 900 node-hours on the Perlmutter supercomputer. For medium-sized organic molecules (up to ∼3k basis functions), the PT2 component adds minimal cost relative to the initial SCF step. This demonstrates that efficient GPU acceleration reduces the practical computational requirements of DHDFT comparable to conventional hybrid DFT. We additionally benchmark a range of LDA, GGA, and MGGA functionals against the revDSD-PBEP86-D4(noFC) isomerization energies. Without dispersion corrections, the SVWN5 LDA functional (MAD 4.47 kJ/mol) outperforms all tested GGAs and MGGAs. With dispersion corrections, only two MGGAs, led by M06-L-D4 (MAD 3.82 kJ/mol), are able to surpass the SVWN5 results.
- 4 May 2026
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 22 (10) , 5060-5072
Double-Hybrid, but Not Double-Cost: GPU-Accelerated DHDFT for the COMPAS-3 Data Set of Polybenzenoid Hydrocarbons
We present the first implementation of double hybrid density functional theory (DHDFT) with all major computational steps accelerated on GPUs. To efficiently utilize GPU hardware, we employ the resolution-of-identity (RI) approximation, transforming electron repulsion integral (ERI) computations into large dense matrix multiplications and enabling basis functions up to g angular momentum. We demonstrate revDSD-PBEP86-D4(noFC)/def2-QZVPP calculations on the entire COMPAS-3x data set of ∼39,000 peri-condensed polybenzenoid hydrocarbon isomer geometries (up to 68 atoms) using just 900 node-hours on the Perlmutter supercomputer. For medium-sized organic molecules (up to ∼3k basis functions), the PT2 component adds minimal cost relative to the initial SCF step. This demonstrates that efficient GPU acceleration reduces the practical computational requirements of DHDFT comparable to conventional hybrid DFT. We additionally benchmark a range of LDA, GGA, and MGGA functionals against the revDSD-PBEP86-D4(noFC) isomerization energies. Without dispersion corrections, the SVWN5 LDA functional (MAD 4.47 kJ/mol) outperforms all tested GGAs and MGGAs. With dispersion corrections, only two MGGAs, led by M06-L-D4 (MAD 3.82 kJ/mol), are able to surpass the SVWN5 results.
- Open Access
A Machine Learning Approach Towards Runtime Optimisation of Matrix Multiplication
The GEneral Matrix Multiplication (GEMM) is one of the essential algorithms in scientific computing. Single-thread GEMM implementations are well-optimised with techniques like blocking and autotuning. However, due to the complexity of modern multi-core shared memory systems, it is challenging to determine the number of threads that minimises the multi-thread GEMM runtime. We present a proof-of-concept approach to building an Architecture and Data-Structure Aware Linear Algebra (ADSALA) software library that uses machine learning to optimise the runtime performance of BLAS routines. More specifically, our method uses a machine learning model on-the-fly to automatically select the optimal number of threads for a given GEMM task based on the collected training data. Test results on two different HPC node architectures, one based on a two-socket Intel Cascade Lake and the other on a two-socket AMD Zen 3, revealed a 25 to 40 per cent speedup compared to traditional GEMM implementations in BLAS when using GEMM of memory usage within 100 MB.
Open AccessA Machine Learning Approach Towards Runtime Optimisation of Matrix Multiplication
The GEneral Matrix Multiplication (GEMM) is one of the essential algorithms in scientific computing. Single-thread GEMM implementations are well-optimised with techniques like blocking and autotuning. However, due to the complexity of modern multi-core shared memory systems, it is challenging to determine the number of threads that minimises the multi-thread GEMM runtime. We present a proof-of-concept approach to building an Architecture and Data-Structure Aware Linear Algebra (ADSALA) software library that uses machine learning to optimise the runtime performance of BLAS routines. More specifically, our method uses a machine learning model on-the-fly to automatically select the optimal number of threads for a given GEMM task based on the collected training data. Test results on two different HPC node architectures, one based on a two-socket Intel Cascade Lake and the other on a two-socket AMD Zen 3, revealed a 25 to 40 per cent speedup compared to traditional GEMM implementations in BLAS when using GEMM of memory usage within 100 MB.
- Open Access
- 17 November 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Machine-Learning Adaptivity, DFT-Level Accuracy, and Semi-Empirical Quantum-Chemistry Speed with Neural-Network Extended Tight-Binding
Fast, general, and interpretable quantum accuracy remains a challenge. To address it, we introduce Neural Network Extended Tight-Binding (NN-xTB), a Hamiltonian-preserving scheme that augments the GFN2-xTB operator with small, bounded, environment-dependent shifts to a compact set of physically interpretable parameters predicted by neural network. NN-xTB confines learning to atom/shell local parameters and to a few global scalings, enhancing local chemical adaptivity while preserving analytic long-range limits, native charge and spin treatment, and self-consistency, and ensuring that all observables are derived from a single adapted electronic state. Across complementary benchmarks, NN-xTB attains DFT-like accuracy at near-xTB cost. On GMTKN55 it achieves a WTMAD-2 of 4 kcal/mol (versus 25 kcal/mol for GFN2-xTB and 9.3 kcal/mol for g-xTB). Relative to strong equivariant machine-learned interatomic potentials (MLIPs), NN-xTB yields the lowest force MAE on eight out of ten rMD17 molecules and delivers 10-40% RMSE reductions on large subsets of MACEOFF23 and SPICE. Under temperature shift on the 3BPA dataset, errors remain substantially below competing MLIPs up to 1200 K, indicating stronger out-of-distribution generalization. On VQM24, the vibrational-frequency MAE drops from 200.6 cm^{-1} (GFN2-xTB) to 12.7 cm^{-1} (more than 90% reduction). The neural component adds less than 20% wall-time overhead, preserving the throughput that motivates semi-empirical quantum methods. By adapting the Hamiltonian rather than directly regressing energies or forces, NN-xTB bridges much of the accuracy gap to DFT while retaining interpretability, correct asymptotics, and robustness, offering a practical route to quantum-accurate molecular simulation and modeling at scale.
- 17 November 2025
- Published by American Chemical Society (ACS) in ChemRxiv
Open AccessMachine-Learning Adaptivity, DFT-Level Accuracy, and Semi-Empirical Quantum-Chemistry Speed with Neural-Network Extended Tight-Binding
Fast, general, and interpretable quantum accuracy remains a challenge. To address it, we introduce Neural Network Extended Tight-Binding (NN-xTB), a Hamiltonian-preserving scheme that augments the GFN2-xTB operator with small, bounded, environment-dependent shifts to a compact set of physically interpretable parameters predicted by neural network. NN-xTB confines learning to atom/shell local parameters and to a few global scalings, enhancing local chemical adaptivity while preserving analytic long-range limits, native charge and spin treatment, and self-consistency, and ensuring that all observables are derived from a single adapted electronic state. Across complementary benchmarks, NN-xTB attains DFT-like accuracy at near-xTB cost. On GMTKN55 it achieves a WTMAD-2 of 4 kcal/mol (versus 25 kcal/mol for GFN2-xTB and 9.3 kcal/mol for g-xTB). Relative to strong equivariant machine-learned interatomic potentials (MLIPs), NN-xTB yields the lowest force MAE on eight out of ten rMD17 molecules and delivers 10-40% RMSE reductions on large subsets of MACEOFF23 and SPICE. Under temperature shift on the 3BPA dataset, errors remain substantially below competing MLIPs up to 1200 K, indicating stronger out-of-distribution generalization. On VQM24, the vibrational-frequency MAE drops from 200.6 cm^{-1} (GFN2-xTB) to 12.7 cm^{-1} (more than 90% reduction). The neural component adds less than 20% wall-time overhead, preserving the throughput that motivates semi-empirical quantum methods. By adapting the Hamiltonian rather than directly regressing energies or forces, NN-xTB bridges much of the accuracy gap to DFT while retaining interpretability, correct asymptotics, and robustness, offering a practical route to quantum-accurate molecular simulation and modeling at scale.
- 8 October 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 21 (20) , 10263-10280
Efficient Algorithms for GPU Accelerated Evaluation of the DFT Exchange-Correlation Functional
Kohn-Sham density functional theory (KS-DFT) has become a cornerstone for studying the electronic structure of molecules and materials. Improving algorithmic efficiency through hardware-aware implementations enables application to larger systems and more efficient generation of larger training data sets for machine-learning. In this work, we present a comparative study of four GPU-accelerated algorithms for evaluating the KS-DFT exchange-correlation (XC) potential with an atom-centered Gaussian basis. Two approaches, both leveraging batched dense linear algebra, are found to outperform the others across a suite of molecular benchmarks. We show that batched formation of the XC matrix from the density matrix yields the best performance for large ( basis functions), sparse systems such as glycine chains and water clusters. In contrast, for smaller and denser systems such as diamond nanoparticles, especially if employing large basis sets, algorithms that use the underlying molecular orbital coefficients offer superior performance, despite their higher formal scaling. Our implementations deliver speedups of 1.4-5.2× for XC potential evaluation relative to leading GPU-accelerated KS-DFT codes, significantly lowering the computational cost and enabling the routine use of larger integration grids. Finally, we outline directions for continued performance improvements in light of emerging GPU architectures with emphasis on utilizing mixed-precision capabilities.- 8 October 2025
- Published by American Chemical Society (ACS) in Journal of Chemical Theory and Computation
- Vol. 21 (20) , 10263-10280
Efficient Algorithms for GPU Accelerated Evaluation of the DFT Exchange-Correlation Functional
Kohn-Sham density functional theory (KS-DFT) has become a cornerstone for studying the electronic structure of molecules and materials. Improving algorithmic efficiency through hardware-aware implementations enables application to larger systems and more efficient generation of larger training data sets for machine-learning. In this work, we present a comparative study of four GPU-accelerated algorithms for evaluating the KS-DFT exchange-correlation (XC) potential with an atom-centered Gaussian basis. Two approaches, both leveraging batched dense linear algebra, are found to outperform the others across a suite of molecular benchmarks. We show that batched formation of the XC matrix from the density matrix yields the best performance for large ( basis functions), sparse systems such as glycine chains and water clusters. In contrast, for smaller and denser systems such as diamond nanoparticles, especially if employing large basis sets, algorithms that use the underlying molecular orbital coefficients offer superior performance, despite their higher formal scaling. Our implementations deliver speedups of 1.4-5.2× for XC potential evaluation relative to leading GPU-accelerated KS-DFT codes, significantly lowering the computational cost and enabling the routine use of larger integration grids. Finally, we outline directions for continued performance improvements in light of emerging GPU architectures with emphasis on utilizing mixed-precision capabilities.