For higher-order derivatives, such as hyperpolarizability, the speed advantage of a HIP-style direct prediction approach would be even larger!
16.10.2025 10:11 — 👍 1 🔁 0 💬 0 📌 0For higher-order derivatives, such as hyperpolarizability, the speed advantage of a HIP-style direct prediction approach would be even larger!
16.10.2025 10:11 — 👍 1 🔁 0 💬 0 📌 0
Thank you!
Cool idea. In principle, other second-order derivatives could similarly be predicted.
The speedup compared to autodiff depends on the dimensionality of the object; for most, it won't be as large as for the Hessian.
DEQuify your force field: More efficient simulations using deep equilibrium models
22.04.2025 09:09 — 👍 0 🔁 0 💬 0 📌 0
ICLR Singapore 🇸🇬
Hit me up to talk about neural samplers and anything to replace molecular dynamics ⚛️
Also I'll talk about how to speedup molecular dynamics using deep equilibrium models (borrowing ideas from DFT)
➡️ AI4Mat Workshop, 28th 11:00 a.m.