Now published in @MLSTjournal!.If you're interested in knowing more about the performance of different MLIPs in terms of both accuracy and efficiency, have a read:.

RT @DynamicsSIAM: "Machine Learning and Data-Driven Methods in Computational Surface and Interface Science" (by Lukas Hörmann, Wojciech G.….

Generate videos in just a few seconds. Try Grok Imagine, free for a limited time.

RT @reinimau: We have a 2-year postdoc position available in the group. Please share widely. The application deadline is September 24th. h….

RT @MLSTjournal: Great new work by @wgstark @IlyesBatatia @reinimau Gábor Csányi et al @compsurfchem @Cambridge_Eng @UNM_Chemistry @Biology….

RT @lbpartay: We are excited to organise the Computational Molecular Science meeting in Warwick again, bringing the modelling community tog….

This is why we need MLIPs that are both highly accurate and computationally efficient. In this work, we show a detailed analysis of the performance of state-of-the-art MLIPs, including ACE, MACE, REANN, and PaiNN, revealing their stronger and weaker sides. Have a good read! 📃🏞️.

Gas-surface systems often require running millions of MD trajectories, to model important dynamical observables, such as reaction probabilities. 🥲Additionally, such properties are highly sensitive to the PES landscape, especially around the reaction barrier.😰⬇️.

⚠️New preprint⚠️"Benchmarking of MLIPs for reactive hydrogen dynamics at metal surfaces" is now available on arXiv: Work with Cas van der Oord, @IlyesBatatia, Yaolong Zhang, Bin Jiang, Gábor Csányi, and @reinimau. ⬇️.

RT @JPhysChem: Adaptive sampling with uncertainty quantification of dynamic reaction probabilities reveals the importance of equivariance a….

Many thanks to all the co-authors: @JWestermayr, @oskar_douglas, James Gardner, Scott Habershon, and @reinimau from @compsurfchem for their incredible help! 🧵[6/6].

We happily share all the important scripts for gas-surface database generation through adaptive sampling (using the NQCDynamics.jl package) in the GitHub repository: Further instructions can be accessed here: .🧵[5/6]

As expected, the equivariant PaiNN models outperform invariant SchNet quite significantly and in every studied aspect. Here, elbow plots generated with SchNet show worrying instabilities, whereas using the same database, PaiNN models provide accurate and smooth PESs. 🧵[4/6]

We show that employing uncertainty quantification (UQ) is necessary to establish the "real" convergence, due to possible high prediction errors, which may cause "false positive" results.🧵[3/6].

We employ adaptive sampling to iteratively train invariant and equivariant (SchNet and PaiNN) models for H2 dissociative adsorption on Cu surfaces, and rather than assessing their performance based only on RMSEs or MAEs, we directly compare reaction probability predictions🧵[2/6]

Finally published (@JPhysChem C)!🚨💫 Are RMSE/MAE metrics enough to assess the performance of ML interatomic potentials? We present adaptively trained MLIPs with a direct assessment of performance through dynamical observables including full UQ. 🧵[1/6].

RT @JWestermayr: 📢 Hi #compchem and #ML community. I have an open PhD position I seek to fill as soon as possible. The focus is on ML for o….

RT @IdilIsmail20: For a gentle introduction to common tools and methods used in computational science, check out our latest preprint now up….

RT @reinimau: We are currently advertising three PhD studentships in my research group (@compsurfchem at @warwickchem ):. PhD project 1:.Co….

RT @ML_Chem: Importance of equivariant features in machine-learning interatomic potentials for reactive chemistry at metal surfaces. (arXiv….

RT @compsurfchem: The M3S Conference ( is in full flow! Some pics from today’s hike around the Lake District ⛰️ fea….