About

Machine learning methods have been increasingly used to model, simulate, and forecast scientific and engineering systems due to their performance and lower computational costs. However, even when a model has a good fit to the data it can be scientifically wrong: creating or destroying mass and energy, producing negative concentrations, violating boundary conditions, or drifting into unphysical regimes over predictions outside the training time horizon. These violations compromise the trust in a learnt model, hindering its adoption in simulation pipelines, digital twins, or decision making.

The problem of incorporating prior knowledge or constraints into these methods has sparked interest in the literature, and different approaches exist: soft penalty terms, projection, correction and differentiable optimisation layers, and hard architectural guarantees. However, several limitations exist when choosing between them, at a cost in accuracy, training stability, and compute.

In this session, we promote discussion across machine learning methods, from new approaches to successful applications in real-world systems. The focus is on promoting transparency, trust, and security in machine learning methods.

Call for Contributions

We invite submissions on all aspects of constrained learning for scientific modelling: methods that enforce conservation laws, symmetries, boundary conditions, stability, or other prior knowledge in machine learning models, theory and benchmarks that compare enforcement strategies, and applications in science and engineering. We welcome work from machine learning, applied mathematics, optimisation, and the natural and engineering sciences.

Submissions are limited to 2 pages (including references) and fall into one of two categories:

1. Ongoing work

   New methods, theoretical results, benchmarks, or applications that are still in progress. Early-stage ideas and negative results that shed light on the trade-offs between accuracy, guarantees, and cost are explicitly welcome. Accepted abstracts are presented as a talk (8 min + 2 min Q&A) and a poster.

2. Tutorials, demos, software libraries & datasets

   Descriptions of interactive tutorials or demos (such as Jupyter notebooks), libraries and packages, or datasets that show how constrained learning methods, architectures, or papers are implemented in practice. Accepted contributions are presented as a 2-minute lightning pitch and a demo during the poster session.

Recently published papers

We also welcome work recently published at top-tier conferences and journals such as JMLR, JAIR, MLJ, PAMI, IJCAI, NeurIPS, ICLR, AISTATS, and ICML. Please state the original venue and upload the PDF of the published paper. Accepted papers are presented as a talk (8 min + 2 min Q&A) and a poster.

Review process

Submissions are reviewed single-blind by the Program Committee, with at least two reviews and one meta-review each. Since the number of talk slots is limited, some contributions may be presented as posters only.

How to submit

Prepare your contribution as a PDF using the PMLR template, and submit it via Microsoft CMT by October 30, 2026 (23:59 AoE). The session is non-archival. If you have any questions, contact the organisers.

Schedule

December 8, 2026. The start time will be announced; times below are relative to the start of the session.

Invited Speaker

The keynote speaker will be announced soon.