NeurIPS 2026 Workshop

XAI4Science: Knowledge Discovery and Trust through Interpretable Foundation Models

December 11–12, 2026 Sydney, Australia Submissions open

Join us in mapping the future of trustworthy AI Foundation Models for Weather and Climate

About

Efficient and Interpretable Foundation Models

Accurate weather and climate forecasting models enable researchers and policymakers to make informed decisions towards a more sustainable future. However, traditional numerical models demand immense computational resources. Machine Learning (ML) has emerged as a viable and efficient alternative, benefiting from the ever-growing volume of collected data. In particular, weather and climate Foundation Models (FMs) achieve comparable results at a fraction of the computational cost, opening the field to smaller countries and research teams with limited resources and representation. Yet while FMs partly address computational costs issues, they introduce a transparency one. Unlike physics-based models, whose predictions can be traced to known equations with a relatively well understood physics, FMs results may prove difficult to understand, especially when they produce physical hallucinations. This opacity may undermine scientific trust and raise equity concerns.

Explainable Artificial Intelligence (XAI) offers a promising path to address this challenge. By revealing which inputs and internal mechanisms drive a model's predictions, XAI methods can help verify that FMs rely on physically meaningful patterns rather than spurious correlations, detect and diagnose physical hallucinations, and communicate the rationale behind forecasts to scientists and decision makers alike. In doing so, XAI can restore the scientific scrutiny that black-box models currently lack, fostering trust in ML-based forecasting and supporting its equitable adoption across the globe. In this workshop we focus on the issues just introduced, by bringing together ML researchers, climate scientists, and policy experts.

Topics & Speakers

What the workshop covers

By bringing together ML researchers, climate scientists, and policy experts, the program focuses on three threads:

  • Ante-hoc interpretabilitySelf-explainable architectures and inductive biases for weather and climate Foundation Models.
  • Post-hoc attribution & evaluationAttribution, probing, and mechanistic interpretability methods, including rigorous evaluation of their faithfulness.
  • Physics-consistent explanationsBenchmarks and methods for validating explanations against known physical laws and causal structure.
Invited Speaker
Gustau Camps-Valls (tentative)

Universitat de València

Invited Speaker
Megan J. Stanley (tentative)

Ellison Institute of Technology Oxford

Invited Speaker
Lily Xu (tentative)

Columbia University

Invited Speaker
David Rolnick (tentative)

McGill University & Mila

Program

Schedule

Tentative — one day, four invited talks, a panel, and two poster sessions.

Time Session
8:00 – 8:15 am Opening Remarks
8:15 – 9:00 am Invited talk I Gustau Camps-Valls Universitat de València
9:00 – 9:45 am Invited talk II Megan J. Stanley Ellison Institute of Technology Oxford
9:45 – 10:30 am Invited talk III Lily Xu Columbia University
10:30 am – 12:00 pm BreakPoster Session I & Coffee Break
12:00 – 1:00 pm Lunch Break
1:00 – 2:00 pm Contributed Talks (I, II, III, and IV)
2:00 – 2:45 pm Invited talk IV David Rolnick McGill University & Mila
2:45 – 4:15 pm BreakPoster Session II & Coffee Break
4:15 – 5:00 pm (Panel) The Interpretability Gap: Is a "Black Box" Prediction Better than No Prediction?
Call for Submissions

Submission guidelines

(1) Regular Paper Track: up to 8 pages, excluding references and appendices.

(2) Tiny Paper Track: up to 5 pages, excluding references and appendices.

Unlimited pages are allowed for references and appendices in the same PDF as the main paper.

Submissions must be in a single PDF file and are required to use the NeurIPS 2026 LaTeX template, available on the NeurIPS 2026 Main Track Handbook.

We welcome optional anonymous submissions of ongoing and unpublished work on any topics related to the workshop.

We require each submission to nominate at least one author to serve as a reviewer, following the NeurIPS 2026 reciprocal review rule, and we aim for each paper to collect at least three reviews.

Contributed talks' speakers will be selected from the top submissions received, giving deserving works a spotlight to a broad audience. Each speaker will have 15 minutes (10+5) for presentation and Q&A.

Important Dates

All dates are in AoE (Anywhere on Earth) time.

Submission Deadline August 29, 2026, AoE
Review Period September 1 – September 15, 2026, AoE
Rebuttal Period Ends September 20, 2026, AoE
Advisory Committee Discussion Ends September 27, 2026, AoE
Notification of Acceptance September 28, 2026, AoE
Workshop Date NeurIPS 2026 (TBD)
Team

Organizers

Leonardo Pesce
Leonardo Pesce

National University of Singapore

Jiawen Wei
Jiawen Wei

National University of Singapore

Max Welling
Max Welling

University of Amsterdam

Gianmarco Mengaldo
Gianmarco Mengaldo

National University of Singapore

Reviewers

Program Committee

Steven Brunton
Wojciech Samek
Ricardo Vinuesa
Nils Thuerey
Gabriele Messori
Xiaoxiang Zhu
Duncan Watson-Parris
Jeff Adie
Wessel Bruinsma
Elizabeth Barnes
Christian Lessig
Rui Mao
Xin Wang
Luwei Xiao