Theory of AI for Scientific Computing

We invite submissions at the interface of learning theory, statistics, numerical methods, algorithm design, and the physical sciences. Submissions may be of any length; in particular, we welcome both short poster abstracts and multi-page papers. The topics of the workshop include (but are not limited to) the following:

Learning-theoretic foundations
• learning-theoretic and statistical analysis of data-driven solutions to important scientific computing tasks such as solving differential equations, inverse problems, sampling, equation discovery, and beyond
• mathematical characterizations of settings in which AI-augmented methods can be expected to improve over traditional (AI-free) scientific computing methods
• end-to-end theoretical studies that consider simultaneously the entire scientific computing pipeline, i.e. not only learning from data but also generating the data itself and integrating the learned model into downstream scientific computing tasks
• non-i.i.d. settings such as active sampling of ground truth solutions, reinforcement learning (RL) of equations and solvers, and transfer learning between different differential equation families, different solution domains, and different initial or boundary conditions
• formalizing concrete goals for scientific discovery
Principled methods for AI-augmented scientific computing
• theoretically-motivated design of neural architectures and loss functions for scientific computing tasks
• mathematical connections between generative modeling and tractable solutions of high-dimensional PDEs
• principled approaches to data-generation, model training, and model deployment
• statistical machinery for certifying the quality and confidence of AI-augmented algorithms and improving their robustness
Connections with other subfields of theory
• sampling
• learning-augmented algorithms (algorithms with predictions) and data-driven algorithm design
• optimization
• randomized numerical linear algebra

Submission criteria:

Papers and abstracts should be submitted as PDF files in any format that has a font size of at least 10 points and margins of at least 1 inch. Submissions are not limited in length, but only the first 8 pages are guaranteed to be reviewed. Accepted submissions will be made public on OpenReview but are non-archival, and we welcome work accepted at previous or upcoming conferences, including COLT 2025 and ICML 2025.

Reviewing and publication:

All submissions will undergo a double-blind peer review process assessing mainly relevance, clarity, and soundness. Reviewing will occur on the OpenReview platform but reviews will not be public. Authors of accepted submissions will be invited to present a poster about it on the day of the workshop (June 30). The organizers will also select up to two submissions to receive best paper and runner-up awards, and their authors will have the opportunity to present short contributed talks.

Submissions can be made through the OpenReview link below; if registering an account there we recommended using an institutional email to avoid an up to two-week moderation period. Please ensure that you follow the submission guidelines outlined in the Call for Papers.

Submission Link

For any questions regarding the submission process, please contact: tasc.organizers@gmail.com

Registration for the TASC workshop is automatically included in your COLT 2025 conference registration fee; no separate registration is required. To register for COLT 2025, go to the COLT 2025 registration website and follow the instructions in the "Registration procedure" section. The early bird registration deadline is May 22.