A Hardware-Native Realisation of Semi-Empirical Electronic Structure Theory on Field-Programmable Gate Arrays

High-throughput quantum-chemical calculations underpin modern molecular modelling, materials discovery, and machine-learning workflows, yet even semi-empirical methods become restrictive when many molecules must be evaluated. Here we report the first hardware-native realisation of semi-empirical electronic structure theory on a field-programmable gate array (FPGA), implementing as a proof of principle Extended Hückel Theory (EHT) and non-self-consistent Density Functional Tight Binding (DFTB0). Our design performs Hamiltonian construction and diagonalisation on the FPGA device through a streaming dataflow, enabling deterministic execution without host intervention. On a mid-range Artix-7 FPGA, the DFTB0 Hamiltonian generator delivers a throughput over fourfold higher than that of a contemporary server-class CPU. Improvements in eigensolver design, memory capacity, and extensions to nuclear gradients and excited states could further expand capability. Combined with the inherent energy efficiency of FPGA dataflow, this work opens a pathway towards sustainable, hardware-native acceleration of electronic-structure simulation and direct hardware implementations of a broad class of methods.

💡 Research Summary

The paper presents the first fully hardware‑native implementation of semi‑empirical electronic‑structure methods on a field‑programmable gate array (FPGA). Using the Xilinx Artix‑7 platform and Vitis High‑Level Synthesis, the authors map both Extended Hückel Theory (EHT) and non‑self‑consistent Density Functional Tight Binding (DFTB0) onto the FPGA fabric, executing the entire workflow—coordinate loading, orbital‑pair generation, Hamiltonian element evaluation, matrix assembly, and eigenvalue diagonalisation—without host intervention.

Key to the design is a streaming data‑flow pipeline. Nested loops over orbital indices are flattened into a flat stream of index pairs generated by a dedicated kernel; each pair is processed element‑wise by downstream kernels. By enforcing an initiation interval of one clock cycle, the pipeline can produce one Hamiltonian element per cycle once filled. The Hamiltonian construction stage is therefore limited only by the speed of the subsequent eigen‑solver. Both methods share the same pipeline interfaces; the only difference lies in the arithmetic of the Hamiltonian element kernel—EHT uses simple overlap integrals and empirical scaling, while DFTB0 combines pre‑tabulated two‑center integrals with Slater‑Koster angular factors.

The eigen‑solver is a cyclic Jacobi algorithm, which dominates runtime and scales roughly as O(N³) with the number of atomic orbitals. Execution times are deterministic for a given geometry; variability arises solely from the number of Jacobi sweeps required for convergence. Benchmarking on linear alkanes from methane to C₁₆H₃₄ shows that the FPGA implementation achieves a throughput more than four times higher than a contemporary server‑class CPU when only the Hamiltonian‑generation kernel is active. When the full workflow is executed, the Jacobi stage becomes the bottleneck, and batching multiple geometries does not significantly reduce per‑geometry time because diagonalisation serialises the pipeline.

Resource utilisation is optimised by employing minimal‑width, arbitrary‑precision data types for indices and addresses, reducing on‑chip BRAM and DSP consumption. A separate “Hamiltonian‑only” configuration replicates the pair‑generation and element‑evaluation kernels to increase parallelism, demonstrating that the diagonalisation stage consumes a substantial fraction of the FPGA resources.

The authors discuss several avenues for future improvement: replacing the Jacobi solver with more efficient QR, Divide‑and‑Conquer, or Lanczos methods; expanding on‑chip memory or integrating high‑bandwidth external memory to handle larger basis sets; adding nuclear‑gradient and excited‑state capabilities; and extending the framework to self‑consistent DFTB variants and correlated ab‑initio methods.

Overall, the work shows that FPGA‑based, streaming, hardware‑native electronic‑structure calculations can deliver high throughput, low energy consumption, and deterministic latency, offering a sustainable path for large‑scale high‑throughput materials screening, machine‑learning‑driven force‑field generation, and other data‑intensive quantum‑chemical workflows.