Accurate and efficient gradients of molecular energy with respect to nuclear degrees of freedom are essential for geometry optimization and molecular dynamics, including simulations that go beyond the Born-Oppenheimer regime. A common approach involves deriving analytical formulas for new electronic structure methods, which is often conceptually difficult and requires tedious coding. Here, we implement analytical, semi-numerical, and automatic differentiation (AD)-based gradient pathways for semiempirical Hamiltonian models in the PYSEQM software package, leveraging both graphics processing unit (GPU) and central processing unit (CPU) architectures. We further extend these capabilities to excited states calculated using the configuration interaction singles and time-dependent Hartree-Fock ansätze. We benchmark wall time, peak memory usage, and accuracy across three molecular families of varying chemical complexity, including systems of up to a thousand atoms. For ground-state simulations, analytical and AD gradients achieve near-identical GPU runtimes, while semi-numerical gradients are slower on GPU but remain competitive on CPU. For excited states, both analytical and custom AD approaches using implicit differentiation show similar performance and low memory requirements, whereas gradients with full AD are memory-limited. AD gradients match analytical ones in accuracy across all tested systems, aided by a quaternion-based diatomic frame rotation for two-center quantities that ensures smooth energy surfaces. Overall, automatic differentiation emerges as a practical alternative to analytical gradients in semiempirical quantum chemistry, offering high accuracy while allowing seamless integration in AI-driven workflows and popular packages, such as PyTorch and JAX. Our results provide actionable guidance for selecting optimal gradient strategies in large-scale ground- and excited-state molecular dynamics simulations.
Accurate and efficient gradients of molecular energy with respect to nuclear degrees of freedom are essential for geometry optimization and molecular dynamics, including simulations that go beyond the Born-Oppenheimer regime. A common approach involves deriving analytical formulas for new electronic structure methods, which is often conceptually difficult and requires tedious coding. Here, we implement analytical, semi-numerical, and automatic differentiation (AD)-based gradient pathways for semiempirical Hamiltonian models in the PYSEQM software package, leveraging both graphics processing unit (GPU) and central processing unit (CPU) architectures. We further extend these capabilities to excited states calculated using the configuration interaction singles and time-dependent Hartree-Fock ansätze. We benchmark wall time, peak memory usage, and accuracy across three molecular families of varying chemical complexity, including systems of up to a thousand atoms. For ground-state simulations, analytical and AD gradients achieve near-identical GPU runtimes, while semi-numerical gradients are slower on GPU but remain competitive on CPU. For excited states, both analytical and custom AD approaches using implicit differentiation show similar performance and low memory requirements, whereas gradients with full AD are memory-limited. AD gradients match analytical ones in accuracy across all tested systems, aided by a quaternion-based diatomic frame rotation for two-center quantities that ensures smooth energy surfaces. Overall, automatic differentiation emerges as a practical alternative to analytical gradients in semiempirical quantum chemistry, offering high accuracy while allowing seamless integration in AI-driven workflows and popular packages, such as PyTorch and JAX. Our results provide actionable guidance for selecting optimal gradient strategies in large-scale ground- and excited-state molecular dynamics simulations.
Automatic differentiation (AD) is a well-established tool that allows us to compute exact derivatives of implemented functions with respect to all their parameters. It does so without requiring any extra coding or knowledge of the analytical derivatives. Derivatives of quantum chemistry algorithms are generally very hard to obtain and often cumbersome to implement. AD shows great potential to solve this problem and to facilitate the implementation of gradients in the context of electronic structure methods. In this dissertation, we have applied AD in a broader context and demonstrate that AD tools can compute arbitrary gradients of quantum chemistry methods with respect to any input parameter and any intermediate function. To this end, we have developed the open-source code DiffiQult, a fully differentiable Hartree-Fock (HF) software package. DiffiQult is a python package that uses AD tools using the external libraries JAX and Algopy. As a proof of concept, DiffiQult is an example of an end-to-end differentiable software package of a quantum chemistry algorithm based solely on AD tools. This application has helped us understand the challenges of AD in quantum chemistry. We outline the details, benefits, and caveats of AD for the HF algorithm in this work. We have chosen HF as a starting point since it contains all the functions included in correlated electronic structure methods. In this way, it opens the door for the development of AD for more accurate algorithms. As an application, we describe DiffiQult within the framework of the fully variational HF method, where we use a gradient-based optimization method to compute system-specific basis functions. As a result, the wavefunction can capture more molecular features, and we obtain a set of basis functions for correlated methods in quantum and classical algorithms. This proves that these functions improve variationally energy while keeping a compact representation in correlated methods by interfacing with the quantum chemistry library PySCF, and with Tequila, a platform for rapid development of quantum algorithms. First, we present optimization results of minimal basis sets with respect to all parameters of the molecules H2O, CH4, NH3, HF. Later we include optimization of the potential energy surface of two isomerization mechanisms of diazene and hydrogens chains. Here, we use a minimal optimized basis set to compute FCI when possible with PySCF and VQE energies of hydrogen chains with Tequila. We obtain similar results compared to a bigger basis set with considerably fewer configurations. Also, we show how this approach improves the estimation of the thermodynamic limit in the hydrogen chain at the HF level. Finally, we apply this approach to the optimization of all atomic orbitals of benzene and compute electronic correlations restricting the active space to the 6 π orbitals of benzene using the UpCCGSD ansatz implemented in Tequila.
Graphical Processing Units (GPUs) are highly parallel, programmable accelerators boasting high peak floating point performance. Over the last couple of years the use of GPUs for general purpose computing have revolutionized quantum chemistry. The computational bottleneck in an ab-initio quantum method is the calculation of a large number of twoelectron integrals. To date, a number of GPU accelerated two-electron integral implementations have been developed significantly improving the performance of a static quantum mechanical (QM) calculation. However, when performing an ab-initio QM gradient calculation, optimization, QM or Hybrid Quantum Mechanical/Molecular Mechanical (QM/MM) dynamics simulation the twoelectron integral derivatives arise as an additional bottleneck. Hybrid QM/MM methods particularly dynamics methods are commonly used to study large chemical/biological systems. These methods are a popular choice used for studying reaction mechanisms, conformational and configurational structures important in glycobiology. Usually a semiempirical QM method is used however, these have shown variable accuracy for the study of carbohydrate conformation and prevents an analytical investigation of electronic structure. The use of a higher level of theory, such as an ab initio method, is desirable however this comes at a much greater computational cost. Together with the bottlenecks above this cost results from the polarization of the QM region within an electrostatic embedding scheme, which requires the calculation of a large number of one-electron integral derivatives. Thus for QM/MM calculations the one-electron integral derivatives becomes a third bottleneck together with the two mentioned above. Recently, the above bottlenecks have become popular GPU acceleration targets. This thesis describes an extension of the GPU based Quantum Supercharger Library (QSL) to perform the above calculations/simulations. In contrast to GPU packages developed from the ground up, the QSL is a library of routines aimed at accelerating legacy codes, such as GAMESS-UK, GAMESS-US and NWChem, used in electronic structure calculations. Algorithms are presented for accelerating the one- and two-electron integral derivatives on a GPU. In addition to the derivatives, the one-electron integral calculation was ported to the GPU in order to remove a small additional cost arising for large QM/MM systems. Furthermore, the use of automatic code generation for generating GPU kernels was explored and compared to the original approach. Using the QSL library implemented in GAMESS-UK several benchmark calculations and simulations were performed. These were performed in double precision on a single GPU (Kepler K20) and compared to a single CPU using the 6-31G basis set. A speedup of up to 9.3X is achieved for an ab initio gradient calculation compared to the CPU running an optimized serial version of GAMESS-UK using the Schlegel method. For a single point QM/MM calculation of cellobiose in different sized water spheres (3267-24843 point charges) speedups of between 13X and 34X is achieved. QSL/GAMESS-UK coupled to the CHARMM molecular dynamics package was then used in order to perform accelerated molecular dynamics simulations. Benchmark QM and QM/MM molecular dynamics simulations were performed on cellobiose in vacuo and in a water sphere (45 QM atoms and 24843 point charges respectively). The QSL is able to perform 9.7 ps/day of ab initio QM dynamics and 6.4 ps/day of QM/MM dynamics. Testing of the auto-generated version of the QSL showed better performance for lower angular momentum classes but reduced performance for higher angular momentum classes. The efficiency of the integral and derivative routines within the QSL library was tested on a computationally intense realistic glycobiological condensed phase free energy computation. Ab-initio pucker free energy surfaces/volumes of !-Ribofuranose and !-Glucopyranose in vacuum and in water were computed. These are the first converged Hartree-Fock/MM free energy simulations for these carbohydrates performed in solution. The value of the ab initio Free energy surfaces/volumes was demonstrated through an analysis of solvent polarization effects that are evident through a comparison of the vacuum and solution minimum free energy pathways. In particular the water structure around the pucker conformers, analysis of the primary alcohol distribution as well as analysis of the electronic structure of these conformers reveals that water significantly affects the free energy pathways, primary alcohol distribution and the barriers for inter-conversion between pucker conformers.
Abstract Automatic differentiation (AD) is a powerful tool that allows calculating derivatives of implemented algorithms with respect to all of their parameters up to machine precision, without the need to explicitly add any additional functions. Thus, AD has great potential in quantum chemistry, where gradients are omnipresent but also difficult to obtain, and researchers typically spend a considerable amount of time finding suitable analytical forms when implementing derivatives. Here, we demonstrate that AD can be used to compute gradients with respect to any parameter throughout a complete quantum chemistry method. We present DiffiQult, a Hartree–Fock implementation, entirely differentiated with the use of AD tools. DiffiQult is a software package written in plain Python with minimal deviation from standard code which illustrates the capability of AD to save human effort and time in implementations of exact gradients in quantum chemistry. We leverage the obtained gradients to optimize the parameters of one-particle basis sets in the context of the floating Gaussian framework.
Semiempirical quantum chemistry (SQC) methods offer fast quantum chemical insights by constructing and solving a parametric effective minimal basis Fock matrix. Establishing suitable parameterizations has long been a challenging and time-consuming task involving tedious grid searches or costly finite-difference gradients of carefully crafted loss functions based on select experimental data. The growing availability of differentiable programming environments that leverage algorithmic differentiation to obtain complex derivatives together with access to a wealth of reliable reference data from ab initio calculations offers a new and more efficient approach. In this work, we extend a previous, basic implementation of SQC methods in PyTorch [J. Chem. Theory Comput. 16, 4951-4962 (2020)] by including global algorithmic considerations into the code design. This allows for improved general applicability and establishes a robust back-end for rapid SQC parameterizations. In particular, we address the general differentiability of the eigensolver and the iterative SCF procedure. The new implementation offers drastic improvements in computing costs as well as memory footprint and provides increased stability in gradient evaluation. We highlight the importance of these advances and their improvements over existing formulations and illustrate their role in the context of SQC parameterization.
The dynamic rearrangement of a molecule at finite temperature is not captured by a single static input geometry to a quantum chemical calculation. Multilevel workflows have established themselves as standard procedures to capture the dependence of the molecule’s configuration on computed properties. Each step within this workflow uses an electronic structure theory method that features the optimal computational time to accuracy ratio. For exploratory steps the computational expense of a method needs to be minimal as several thousand energy and nuclear gradient evaluations are needed. Semiempirical tight binding methods, such as GFN2-xTB are often used for such purposes. However, these methods usually lack the accuracy to rank conformers relative to each other within chemical accuracy (2 kcal/mol). Therefore, so-called “low-cost” DFT methods are commonly used for geometry optimizations and conformer ranking. The combination of (semi)empirical correction models with small basis sets leads to good accuracy at moderate computational cost. However, even these highly efficient methods reach the limits of feasibility for very large systems (> 200 atoms). This work proposes solutions to partially alleviate those restrictions and allow routine calculations of these system sizes outside of large high-performance computing facilities. The key to unlock those calculations is to combine tailored algorithmic developments with consumer-grade GPUs. A GPU-enabled linear algebra library is developed to accelerate further development and make the entry to GPU-accelerated software development easier. To demonstrate the ease of implementation, the addition of Mulliken-approximated exchange to GFN2-xTB will be shown. In most semiempirical methods, the general eigenvalue solver algorithm becomes the computational bottleneck, dominating the walltime with increasing system size. A density matrix purification algorithm is tailored to give up to six-fold acceleration of GFN2-xTB calculations using a mixed-precision approach. Lastly, 3c-methods are added to the GPU-accelerated TeraChem software package. In combination with the “mixed-precision” scheme for computing the electron repulsion integrals, this addition decreases the runtime of DFT calculations by up to an order of magnitude on mid-level consumer hardware. The overall goal of these projects is to make GPU-accelerated calculations in quantum chemistry more available. This achieve by exposing algorithmic developments that are performant on reasonably priced hardware and making GPU computing more accessible to non-experts.
We present a GPU-accelerated software package for the evaluation of analytic two-electron energy and gradient contributions in Hartree-Fock (HF) and Density Functional theory (DFT) calculations. The implementation is provided as a Python library with a C++ backend, enabling straightforward integration into modern computational chemistry and drug-discovery workflows. The code supports single-point energy and nuclear gradient evaluations on both single- and multi-GPU systems, and employs MPI-based parallelization with dynamic load balancing in multi-node environments. We report comprehensive benchmarks demonstrating favorable scaling with respect to system size, as well as high throughput for batched evaluations relevant to molecular dynamics, geometry optimization, and large-scale virtual screening. Parallel execution of a single system was carried out on up to 24 A100 GPUs. The implementation builds on optimized GPU-enabled variants of the LibintX and GauXC libraries to efficiently compute density-fitted Coulomb, semi-numerical exchange, and exchange--correlation contributions.