Senior Researcher in neuro-symbolic AI and scientific machine learning
I build neural networks that reason with modal logic, one-shot PDE solvers, and AI agents that run experiments on particle accelerators.
I am Antonin Sulc (Antonín Šulc), a Senior Researcher at Innovatrics, and I remain affiliated with Lawrence Berkeley National Laboratory (LBNL) as a Researcher.
Previously a Senior Data Scientist at the Helmholtz Association (DESY/XFEL), I pioneered the use of domain-specific language models in scientific contexts. Holding a Ph.D. in Computer Vision from the University of Konstanz, I actively apply the mathematical rigor of high-dimensional signal processing to extract actionable insights from complex, multi-modal data.
My current work centres on Modal Logical Neural Networks, a family of models that make modal logic (necessity, possibility, knowledge, belief, obligation) differentiable, with applications to multi-agent systems, quantum circuits and finance. The open-source library torchmodal implements them for PyTorch.
Read a short interview about my work and background: 3Q4 with Antonin Sulc (ATAP, LBNL).
A scikit-learn style Python library providing a geometric approach to learning composable, highly interpretable representations of discrete event sequences. Published in PMLR Vol. 282 (NeurReps 2025).
A framework introducing Differentiable Modal Logic for PyTorch, enabling the training of Modal Logical Neural Networks (MLNNs).
A lightweight framework for solving PDEs in one shot via Fourier features with exact analytical derivatives. It bypasses automatic differentiation and iterative training loops for linear and nonlinear problems.
Showing 22 of 22 publications
Published in PMLR Vol. 282, NeurIPS 2025 Workshop on Symmetry and Geometry in Neural Representations (NeurReps)
PMLR 282:492–512, 2026
Learns event-sequence embeddings in which a sequence is the sum of its events, with a hyperbolic variant for hierarchical data.
arXiv:2603.04019 (2026)
Modal logic on continuous state spaces through neural SDEs, with logic-informed neural networks as the logical counterpart of PINNs.
Presented at ICLR 2026 Workshop on Financial AI (FinAI)
arXiv:2603.12487 (2026)
MLNNs as a differentiable logic layer for finance: regulatory guardrails, stress testing and collusion detection.
Oral talk 🎤 at 3rd International Conference on Neuro-Symbolic Systems (NeuS 2026), University of Southern California
arXiv:2602.12083 (2026), NeuS 2026 proceedings to follow
A hands-on tutorial on learning trust, causality and permissions in multi-agent AI systems from behaviour alone.
Presented at 📐 ICLR 2026 Workshop on AI & PDE
arXiv:2602.10541 (2026)
Solves linear PDEs with a single least-squares call, using sinusoidal random features whose derivatives are exact and cheap. No autodiff and no training loop.
arXiv:2602.08880 (2026)
Treats quantum circuit design as differentiable logic programming: learnable gate switches are optimised to satisfy logical axioms.
Oral talk 🎤, presented at 🎉 20th International Conference on Neurosymbolic Learning and Reasoning (NeSy 2026), Lisbon
arXiv:2512.03491 (2025), NeSy 2026 proceedings to follow
Neural networks that reason about necessity and possibility across possible worlds, with differentiable Kripke semantics and a learnable accessibility relation.
Physical Review Research 8, L012017 (2026)
A language-model agent that plans and runs multistage physics experiments at the Advanced Light Source, cutting preparation time about a hundredfold.
APL Machine Learning 4, 016103 (2026)
A framework for running agentic AI safely in the control rooms of large scientific facilities, first deployed at the Advanced Light Source.
NeurIPS ML4Physics Workshop, 2025
Physics-informed neural networks that infer a quantum system's noise parameters from sparse measurements via the Lindblad equation.
16th International Particle Accelerator Conference (IPAC'25)
Retrieval-augmented generation for the electronic logbooks of Fermilab, Jefferson Lab, LBNL and SLAC.
EPS-High Energy Physics, 2025
Probes what open LLMs such as Llama, Qwen and Gemma actually encode about quantum chromodynamics.
42nd International Conference on High Energy Physics (ICHEP 2024)
A generative language model built on a corpus of quantum chromodynamics literature.
NeurIPS ML4Physics Workshop, 2024
A vision for decentralised, LLM-powered multi-agent control of particle accelerators.
15th International Particle Accelerator Conference (IPAC'24)
A multi-laboratory study of retrieval-augmented generation for making accelerator logbooks searchable and useful.
NeurIPS ML4Physics Workshop, 2023
A language model fine-tuned on public accelerator literature, with training questions generated automatically.
IET Information Security, 2023
Unsupervised anomaly detection for European XFEL control-system logs using word embeddings and hidden Markov models.
Journal of Physics: Conference Series 2420, 012070 (2023)
Data-driven anomaly detection for the superconducting RF cavities of the European XFEL.
Oral 🚀 at BMVC 2021
British Machine Vision Conference (BMVC), 2021
Recovers the shape of a refractive surface from a single image, using an energy function built on Snell's law.
Journal of the Optical Society of America A 38(8), 2021
Journal extension of depth estimation from wavelength-dependent defocus blur.
IEEE International Conference on Image Processing (ICIP), 2019
Estimates depth from the way defocus blur differs across spectral bands.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016
Sparse coding of light fields with disparity-aware dictionaries reveals depth and multi-layer scene structure.
Modal Logic Neural Networks
Differentiable Modal Logic for Multi-Agent Diagnosis, Orchestration and Communication
FastLSQ: Solving PDEs in One Shot
Modal Logical Neural Networks for Financial AI
Towards Monocular Shape from Refraction
For collaborations, invited talks or press enquiries, email me or connect on LinkedIn.
