Vaidotas (Vaidas) Šimkus

I'm a Senior ML researcher and engineer who values simple, scalable solutions to hard problems affecting innovation. At Orbital, I build atomistic foundation models and post-train LLMs for accelerated materials discovery and hardware engineering, applied to carbon capture and data center cooling. I'm also the core maintainer of Orb, our open-source SoTA forcefield model.

I did my PhD in Machine Learning at the University of Edinburgh, advised by Michael Gutmann, focusing on probabilistic modelling with missing data. I also hold a BEng in Software Engineering from the University of Southampton.

Email / GitHub / Twitter / LinkedIn

Papers

Mofasa: A Step Change in Metal-Organic Framework Generation

Vaidotas Šimkus, Anders Christensen, Steven Bennett, Ian Johnson, Mark Neumann, James Gin, Jonathan Godwin, Benjamin Rhodes

arXiv, 2025
bib / arxiv / website

We introduce Mofasa, an all-atom generative model with SoTA generation fidelity for Metal-Organic Frameworks, scaling to crystal systems with up to 500 atoms per unit cell.

Orb-v3: Atomistic Simulation at Scale

Ben Rhodes, Sander Vandenhaute, Vaidotas Šimkus, James Gin, Jonathan Godwin, Tim Duignan, Mark Neumann

arXiv, 2025
bib / arxiv / code / blog / blog2 / matbench

We introduce Orb-v3, a family of universal interatomic potentials, that offer near-SoTA performance across various evaluations with a >10x reduction in latency and >8x reduction in memory.

CFMI: Flow Matching for Missing Data Imputation

Vaidotas Šimkus, Michael Gutmann

arXiv, 2025
bib / arxiv / code

We introduce a conditional flow-matching method for missing data imputation, which matches or outperforms a wide range of existing methods on tabular and time-series data.

Improving Variational Autoencoder Estimation from Incomplete Data with Mixture Variational Families

Vaidotas Šimkus, Michael Gutmann

Transactions on Machine Learning Research (TMLR), 2024
url / bib / arxiv / code / poster / dmlr-at-iclr2024

We show that missing data increases the complexity of the posterior distribution of the latent variables in VAEs. To mitigate the increased posterior complexity we introduce two strategies based on (i) finite and (ii) imputation-based variational-mixtures.

Conditional Sampling of Variational Autoencoders via Iterated Approximate Ancestral Sampling

Vaidotas Šimkus, Michael Gutmann

Transactions on Machine Learning Research (TMLR), 2023
url / bib / arxiv / code

We link a structured latent space in VAEs, a commonly desired property, to poor conditional sampling performance of Metropolis-within-Gibbs (MWG). To mitigate the issues of MWG we introduce two original methods for conditional sampling of VAEs: AC-MWG and LAIR.

Variational Gibbs Inference for Statistical Model Estimation from Incomplete Data

Vaidotas Šimkus, Ben Rhodes, Michael Gutmann

Journal of Machine Learning Research (JMLR), 2023
url / bib / arxiv / code / poster / slides / jmlr-to-neurips2023 / demo

We propose a new method for statistical model estimation from incomplete data, called variational Gibbs inference (VGI). Whilst being general-pupose, the proposed method outperforms existing VAE and normalising flow specific methods.

Learning Job Titles Similarity from Noisy Skill Labels

Rabih Zbib, Lucas Alvarez Lacasa, Federico Retyk, Rus Poves, Juan Aizpuru, Hermenegildo Fabregat, Vaidotas Šimkus, Emilia García-Casademont

FEAST workshop at ECML-PKDD, 2022
url / bib / arxiv / dataset

We propose an unsupervised representation learning method for a job title similarity model using noisy skill labels. We show that it is highly effective for tasks such as text ranking and job normalization.