Towards black-box hyperparameter optimization of ONCV pseudopotentials

by Dr Austin Zadoks (PSI/LMS)

Wednesday 5 Nov 2025, 15:00 → 16:00 Europe/Zurich

OVGA/200

The pseudopotential approximation is essential to the tractability of the plane-wave DFT calculations underpinning much of ab-initio computational materials science. Of the available formalisms for constructing pseudopotentials, the norm-conserving (NC) prescription remains the most practical in terms of simplicity of implementation, particularly with regards to the computation of forces and stresses, the derivatives required for density functional perturbation theory, and the various quantities required for beyond-DFT theories. The state-of-the-art in soft NC pseudopotentials lies in the optimized norm-conserving Vanderbilt pseudopotential (ONCV) method [1] implemented in the `oncvpsp` code by D. R. Hamann. [2] Various manually-tuned tables of ONCV pseudopotentials, such as the SG15, GBRV, and PseudoDojo, have seen wide adoption in the field and even shown best-in-class accuracy with respect to all-electron reference results. [3] However, as more all-electron reference data are made available, and new exchange-correlation functionals such as SCAN gain popularity, it has become clear that automated procedures to optimize these potentials are required. Shojaei and coworkers [4] have shown that modern multi-objective evolutionary algorithms are able to efficiently perform ONCV pseudopotential generation, discovering Pareto-optimal combinations of softness and accuracy without human intervention. In this seminar, I present our proposed extension of the optimization problem to multiple fidelities, leveraging Bayesian optimization tools developed for optimizing the hyperparameters of deep neural networks [5], more varied all-electron reference data [3], and recent code developments for generating meta-GGA ONCV pseudopotentials [6].

 

[1] D. R. Hamman, Phys. Rev. B. 88, 085117 (2013)

[2] https://github.com/oncvpsp/oncvpsp

[3] E. Bosoni, et al. Nat. Rev. Phys. 6, 45-58 (2024)

[4] M. F. Shojaei, et al. Comp. Phys. Comms. 283, 108594 (2023)

[5] M. Olson, et al. AutoML 2025. ABCD Track. (2025)

[6] “METAPSP-1.0.1 (alpha release)” D. R. Hamann. http://mat-simresearch.com/