Differentiable Systems and Scientific Machine Learning Workshop

About

Automatic differentiation is a key technology for most machine learning models and inverse problems, including surrogate models that simulate + optimize complex scientific phenomena. But can we go beyond individual models?

More precisely, can we build entire differentiable systems that combine multiple components such as adjoint-based simulators, mathematical solvers, surrogate models, and 3D renderers to tackle real scientific challenges?

This workshop aims to answer that question by bringing together experts from around the community. We expect contributions that present advances in:

Novel approaches to scientific machine learning (SciML).
Design, implementation, and analysis of end-to-end differentiable systems.
Challenges and solutions when propagating gradients and performing optimization through complex operations and pipelines.
Demonstrations of differentiable systems specifically, or SciML broadly, on real-world scientific problems.

(See Call for Papers for more details)

Important dates

Submission deadline: 10 October 2025, AoE
Notification: 31 October 2025
Workshop: 6 or 7 December 2025

Tentative Schedule

All times below are GMT+2.

Time	Event
09:00 - 09:15	Opening remarks
09:15 - 10:00	Invited Talk 1
10:00 - 10:45	Invited Talk 2
10:45 - 11:00	Coffee Break
11:00 - 12:00	Contributed Talks (selected from submissions)
12:00 - 13:00	Panel Discussion
13:00 - 14:00	Lunch Break
14:00 - 14:45	Invited Talk 3
14:45 - 15:30	Invited Talk 4
15:30 - 15:45	Coffee Break
15:45 - 16:45	Contributed Talks (selected from submissions)
16:45 - 18:00	Poster Session

Invited Speakers

Petros Koumoutsakos

Professor of Computing in Science and Engineering, Harvard University, and leading expert in AI for science.

Patrick Kidger

Core maintainer of the JAX scientific software ecosystem and technical lead at Cradle Bio.

Nils Thuerey

Professor of Computer Science at the Technical University of Munich, working on physics-based simulation and machine learning, and pioneer in differentiable physics research.

Astrid Walle

Siemens Energy, expert in ML + data science for real-world simulation problems.

Call for Papers

We invite previously unpublished submissions in a form of short papers, including those describing work in progress, as long as they represent advances within one or more of the following topics:

Challenges and opportunities in end-to-end differentiable systems.
Benchmarks involving SciML components or differentiable systems.
Differentiable versions of traditionally non-differentiable operations.
Theory and practice of automatic differentiation and differentiable programming.
Theoretical insight into properties of differentiable systems.
Differentiable programming in heterogeneous or distributed computing environments.
Surrogate modelling (physics-infused methods, neural operators, ...).
Foundation models for simulation / science.
Hybrid models combining mechanistic and data-driven components.
Gradient-based optimization in SciML, e.g. for inverse problems and parameter estimation.
Alternative methods of performing optimization (such as Bayesian optimization).
Reinforcement learning and optimal control.
Case studies applying SciML / differentiable systems to a scientific domain (e.g., weather / climate, physics, engineering design).

Submission Guidelines

Length: up to 4 pages plus unlimited references
Format: NeurIPS 2025 template
Reviews: double-blind, please anonymize your submission
Proceedings: no formal proceedings, accepted papers will be shared on the website

Submission website: Submit on OpenReview