Matthias Rupp

Machine Learning for Physics, Chemistry, and Materials

[Matthias Rupp @_MatthiasRupp]

New work by T. Bischoff, B. Jäckl, and me:

Benchmarking machine-learning potentials for molecular dynamics requires running accelerated simulations. We provide a fully automated benchmark for this.

Preprint: https://t.co/6qgUmulKOB
Repository: https://t.co/GCYpYSLLtj

[Matthias Rupp @_MatthiasRupp]

Ph.D. position in molecular dynamics simulations for ion-beam-based nanoanalytics at LIST

Secondary ion mass spectrometry (SIMS), molecular dynamics simulations, machine-learning interatomic potentials, programming.

Position M-2422 at https://t.co/g8C7AerGO8

[Matthias Rupp @_MatthiasRupp]

Work by Marcel F. Langer, Florian Knoop, Christian Carbogno, Matthias Scheffler, and me:

Accurate thermal transport via Green-Kubo and message-passing neural networks, demonstrated for zirconia.

https://t.co/W52Jg6O6qU

Data and code available.

[Matthias Rupp @_MatthiasRupp]

Work by Stephen R. Xie, Richard Hennig, and me:

Ultra-fast potentials use explicit, effective many-body terms learned from data. They are data-efficient, often accurate enough, interpretable, and ...very fast.

https://t.co/jphnieIITL

Code at https://t.co/pHGtTE3IIV

[marcel ⊙ @marceldotsci]

🚀 New work: "Stress and heat flux via automatic differentiation" w/ @frankthorben and @flokno_phys. We give an overview of using AD for stress and heat flux, and provide an implementation in JAX. See https://t.co/A1omuOgLUV; code at https://t.co/JqtIl7HngX. ���⤵️

[Matthias Rupp @_MatthiasRupp]

I asked ChatGPT for a reference for a claim, and it provided one: It was nicely formatted, but the author list and title were from one publication, whereas year, journal, and DOI were from another, neither related to the claim :-)

[Matthias Rupp @_MatthiasRupp]

Post-doctoral researcher position on Machine Learning of Plasma-Assisted Deposition Processes available in my group.

https://t.co/a5SHINCmJi

[Matthias Rupp @_MatthiasRupp]

ELLIS Ph.D. Program: Call for applications 2022 is open
https://t.co/hLn48rCl9B
Deadline November 15

[Matthias Rupp @_MatthiasRupp]

Specifically (in the order they were published):

SF, BS, CM, SOAP, OMF, BoB, WST, MTP, MBTR, HDAD, DECAF, FCHL, IDMBR, MOB, ACE, NICE, GM, MILAD.

[Matthias Rupp @_MatthiasRupp]

Work by Marcel Langer, Alex Goeßmann, and me:

We review and discuss current representations and relations between them, and benchmark some of them.

https://t.co/sHdEyuOnPB

We cover all 18 representations we were aware of at the time of writing.

[Matthias Rupp @_MatthiasRupp]

Publishing scientific code for atomistic systems:
The Journal of Chemical Physics has a dedicated section for "Chemical Physics Software" since last year.
https://t.co/gtQzVmPQWc

[Matthias Rupp @_MatthiasRupp]

For predictive uncertainty, the convex hull definition of interpolation might thus not be a suitable choice, and Ref. 1 proposes an ad-hoc measure of training data density around a new data point as an indicator for predictive accuracy. 3/3

[Matthias Rupp @_MatthiasRupp]

Interesting work by others:

[1] https://t.co/FwrNdAhrWx
[2] https://t.co/s9JowPocML

They show that in high-dimensional feature spaces, new data almost always lies outside the convex hull of the training data. Machine learning models thus extrapolate in this technical sense. 1/3

[Matthias Rupp @_MatthiasRupp]

In this, it seems the dimensionality of the data within the embedding space (that is, the dimension of the smallest affine subspace containing the data, or, equivalently, the dimension of the convex hull of the data) seems to be the relevant quantity (Fig. 1 in Ref. 2). 2/3

[Matthias Rupp @_MatthiasRupp]

@cetinkayakoc Is there a publication from Chui that you would particularly recommend? We are approximating high-dimensional noise-free functions (and their derivatives). Regarding B-splines, we mostly follow de Boor's Practical Guide to Splines textbook.

[Matthias Rupp @_MatthiasRupp]

New work by S. Xie, R. Hennig and me:

We combine effective low-order many-body expansions with splines and curvature regularization for machine-learning potentials that are robust, interpretable, accurate, and as fast as classical empirical potentials.

https://t.co/ucCn78ge7f

[Matthias Rupp @_MatthiasRupp]

@cetinkayakoc B-splines do have a long history of usefulness in different application areas. I assume FHE means fully homomorphic encryption in your context? We chose them for compact support, smoothness, robustness (fewer oscillations, e.g., Runge's phenomenon), and their derivatives.

[Matthias Rupp @_MatthiasRupp]

@davkovacs10 Yes, that would be interesting. The plots in the two manuscripts are not directly comparable (different losses, training set sizes, cut-off radii, and systems).

[Matthias Rupp @_MatthiasRupp]

@simonbatzner Not yet, but we have some promising preliminary results on multi-component systems. I would expect the comparatively low complexity (few coefficients) of the model to favor speed and data efficiency but limit accuracy. We want to empirically investigate these trade-offs further.

[Matthias Rupp @_MatthiasRupp]

Review of applications begins immediately until the position is filled. Please send your curriculum vitae and a short statement on why you apply for this particular position to [email protected].

[Matthias Rupp @_MatthiasRupp]

Postdoc position on

Machine Learning for Quantum Monte Carlo

An exciting opportunity for research on data-driven electronic structure calculations for materials.
two years, @UniKonstanz, @trex_eu, group: https://t.co/qtlRcPGAny

[Matthias Rupp @_MatthiasRupp]

Profile: ideally knowledgeable and experienced in quantum Monte Carlo and machine learning, with a strong background in mathematics and programming; experience with the AiiDA framework is a plus