On the practical applicability of modern DFT functionals for chemical computations. Case study of DM21 applicability for geometry optimization

TL;DR

The paper investigates the practical applicability of the neural-network DM21 exchange-correlation functional for geometry optimization within DFT, emphasizing oscillations in gradients and the lack of analytic nuclear gradients as key challenges. By implementing DM21 in PySCF with GeomeTRIC/TRIC coordinates and using numerical gradients, the study quantitatively compares DM21 against PBE0 and SCAN across standard benchmarks, including LMGB35 and glycine conformers. The results show DM21 delivers high energy accuracy but does not consistently improve optimized geometries and is slower due to SCF convergence issues, suggesting current neural-network functionals may require training that accounts for PES curvature and derivative behavior. The work proposes strategies to extend practical applicability, such as incorporating PES curvature information into training and rethinking how gradients are treated, to enable DM21-like functionals to reliably model new substances. This has significant implications for the adoption of NN-based XC functionals in routine geometry optimizations and materials/chemical discovery workflows.

Abstract

Density functional theory (DFT) is probably the most promising approach for quantum chemistry calculations considering its good balance between calculations precision and speed. In recent years, several neural network-based functionals have been developed for exchange-correlation energy approximation in DFT, DM21 developed by Google Deepmind being the most notable between them. This study focuses on evaluating the efficiency of DM21 functional in predicting molecular geometries, with a focus on the influence of oscillatory behavior in neural network exchange-correlation functionals. We implemented geometry optimization in PySCF for the DM21 functional in geometry optimization problem, compared its performance with traditional functionals, and tested it on various benchmarks. Our findings reveal both the potential and the current challenges of using neural network functionals for geometry optimization in DFT. We propose a solution extending the practical applicability of such functionals and allowing to model new substances with their help.

Figures (5)

• Figure 1: Schematic diagram of the DFT geometric optimisation algorithm
• Figure 2: PES of H$_{2}$ molecule with a shortened bond length (left) in minimum (right) calculated with DM21, SCAN, and SCAN with noise in $e_{xc}$ and $v_{xc}$. All the points were obtained from converged SCF single point calculations with energy tolerance of $10^{-11} E_{h}$ (Hartree energy).
• Figure 3: Time tests of PBE0, DM21 and CCSD(T). Horizontal axis shows the computation time of the above systems using CCSD(T) method, while vertical axis shows DFT computation time of these systems with DM21 and PBE0 functionals. Calculation have performed at Google Colab machine with 2 CPU cores of Intel(R) Xeon(R) CPU @ 2.20GHz and 13GB RAM.
• Figure 4: Comparison of DM21 and analytical functionals on the LMGB35 benchmark
• Figure 5: Benchmark of high-level optimized geometries