Arghya Bhowmik

Research profile

Research summary

Arghya Bhowmik is a computational materials scientist at the Technical University of Denmark working at the junction of density functional theory (DFT), machine-learned interatomic potentials, and electrochemical energy materials. His OpenAlex-resolved record (h-index 29, i10-index 54) shows a sharply accelerating curve: countsByYear climbs from 6 works in 2019 to 20 in 2021, peaks at 39 in 2022, and holds at 30-33 through 2024, with 69 works over the most recent three years. Topic mass concentrates in machine learning in materials science (121 works), advanced chemical physics (79), and battery materials (60), with computational drug discovery (42) as a smaller strand. Two 2026 arXiv papers reviewed by Evolve define the current program: SoLiD26, a 15.4-million-structure first-principles solid-liquid interface dataset that cuts MACE force errors roughly 3x against a pretrained baseline, and the first reference-free neural electronic initializer (valence, PAW augmentation and spin aware) for PAW DFT, which accelerates self-consistent field convergence. Evolve rates both at medium reproducibility readiness, noting that SoLiD26's data remains request-only and the initializer's headline gains are out-of-distribution best cases. Co-authorship clusters tightly around a DTU-anchored group of 12 distinct recent collaborators.

Research keywords

Highlights

• SoLiD26: 15.4M first-principles solid-liquid interface structures for ML potentials

  The 2026 arXiv dataset paper assembles 15.4 million DFT solid-liquid interface structures; training MACE on it cuts force errors roughly 3x versus a pretrained baseline. Evolve calls it valuable but notes the data remains request-only.

• First reference-free neural electronic initialization for PAW DFT

  The companion 2026 paper introduces a valence-, PAW-augmentation- and spin-aware charge density initializer that accelerates materials DFT. Evolve credits credible ablations but cautions that the headline gains are out-of-distribution best cases with thin statistics.

• A two-track program: large interface datasets plus faster DFT itself

  Read together, the reviewed work pairs dataset construction for machine-learned potentials with algorithms that cut self-consistent field cost, matching the topic profile of machine learning in materials science (121 works) and battery materials (60 works).

Timeline

1. 2010Earliest records in the countsByYear series appear, with 2 works in 2010 and 1 in 2011 — the low pre-surge baseline of a slow-starting record.
2. 20196 works recorded — the last pre-surge year of modest output before activity multiplies.
3. 2021Output jumps to 20 works, beginning the sustained post-2021 surge.
4. 2022countsByYear peaks at 39 works; 2023 and 2024 hold near the peak at 30 and 33.
5. 2026Two arXiv papers reviewed by Evolve appear: the SoLiD26 solid-liquid interface dataset and reference-free neural electronic initialization for PAW DFT; the recent three-year window totals 69 works.

Papers in the Evolve catalog

• 2026-09-23

  SoLiD26: A First Principles Solid-Liquid Interface Dataset for Machine-learned Interatomic Potentials

• 2026-09-18

  Complete Neural Electronic Initialization Accelerates Materials DFT

Frequent co-authors