AI Methods for Science

About the Course

AI methods are increasingly central to how science gets done, spanning simulation, experiment, theory, and observation. This course aims to equip students with the methods to understand and carry out research at the intersection of AI and the natural sciences. Topics include probabilistic inference, neural networks that encode physical symmetries and domain knowledge, generative models for scientific data, and simulation-based inference. While framed in terms of scientific applications, the methods discussed extend well beyond scientific research, with broad applicability across industry and general AI R&D.

A major focus of the course is on large language models and their emerging role in science. As LLMs become more capable of scientific reasoning and operating autonomously, understanding how to evaluate, adapt, and collaborate with these systems is becoming essential. We explore what it means to work alongside AI scientists, and how to critically assess their capabilities as well as limitations.

Applications are drawn from domains including physics, materials science, and biology. The course involves two assignments emphasizing method design and critical analysis in collaboration with AI tools, plus two projects: a midterm applying AI methods to a scientific problem, and a final project finetuning an LLM to elicit a scientific capability.

Learning Objectives

By the end of this course, students will be able to:

• Apply probabilistic inference and sampling methods (e.g., MCMC) to scientific problems
• Design neural networks that encode scientific domain knowledge
• Train generative models (e.g., diffusion models) to emulate scientific data distributions
• Use simulation-based inference to connect simulators with observations
• Evaluate the edges and limitations of LLM capabilities for scientific reasoning
• Develop intuitions for how to collaborate effectively with AI systems on research tasks
• Read and understand AI-for-science research papers

Logistics

Lecture

Mon/Wed 12:20–1:35pm, CAS 218

Discussion

Tue 11:15am–12:05pm, MUG 205

Instructor

Siddharth Mishra-Sharma (smishras@bu.edu)

TF

Wanli Cheng (cwl1997@bu.edu)

Office Hours

Tue 3–5pm or by appointment, CDS 1528

TF Office Hours

Mon/Wed 2:15–3:15pm or by appointment, CDS 14th floor Green Corner

Resources

• Syllabus: PDF
• Discussion: Ed Discussion
• Assignment/lab submission: GitHub Classroom
• Computing: GPU access through the Shared Computing Cluster (SCC) and LLM API/finetuning credits will be provided for project work

There is no required textbook. Many readings reference Understanding Deep Learning by Simon J.D. Prince (MIT Press, 2023); the PDF is available on the website. Other readings are drawn from research papers and online resources.

Schedule and Optional Reading

This schedule will change as the course progresses.

Discussion Sections

Tuesdays 11:15am–12:05pm in MUG 205.

Topics Not Covered

Due to time constraints, this course does not cover several areas in AI for science, including neural operators, physics-informed learning, surrogate modeling, causal inference, interpretability methods, experimental design, active learning, and recent AI-for-math developments (e.g., LLM-guided theorem proving). Some of these may be covered in later weeks as the course evolves.

Assessment

Discussion Labs

Weekly in-class labs reinforce lecture material through hands-on programming. Students work through a notebook during discussion, exploring implementations and comparing results. Graded on participation and completion. Labs are due end of day Wednesday.

Assignments

Two assignments develop skills in method design and critical analysis. AI tools may be used freely, but the analysis and interpretation require critically engaging with what was produced. The discussion labs build foundational skills for these assignments.

• Assignment 1: Sampler Synthesis: Design and stress-test a novel sampling or variational inference method
• Assignment 2: Ablation Archaeology: Systematically ablate components of a geometric neural network to understand what each design choice contributes and why

Midterm Project

Teams of 2–3 conduct a mini research project applying methods from the first half of the course (inference, architectures, generative models) to a scientific problem. Choose from a suggested list or propose your own. Deliverable is a ~4 page workshop-style paper + code.

Final Project

Teams of 2–3 identify a scientific capability that current large language models struggle with, then finetune a language model to improve that capability. This is a two-stage project:

• Proposal: Demonstrate that LLMs struggle at a specific scientific capability by curating a dataset or benchmark
• Final report (~6 pages, NeurIPS format): Fine-tuned model and evaluation showing improvement, including artifacts used for fine-tuning (datasets, code, reinforcement learning environments)

Timeline

Policies

Attendance

Regular attendance in lectures is expected. Please notify the instructor of planned absences.

Late Work

Late submissions are not accepted without prior arrangement. Extensions may be granted for documented emergencies.

Collaboration

Discussion of concepts and approaches is encouraged. However, all submitted code and written work must be your own. When collaborating, you must acknowledge your collaborators.

AI Tools

Learning to work effectively with AI is itself a course objective. Use AI tools freely to explore ideas, debug code, and deepen understanding. Focus on building genuine competence—understanding why something works, not just that it works. Disclose AI assistance in submissions, including its form and extent. See also the CDS GAIA policy.

Academic Conduct

All students are expected to read and abide by the BU Academic Code of Conduct. Plagiarism includes copying or restating work or ideas of another person or AI software without citing the source. In computing coursework, this includes sharing code, reusing code across courses without permission, and uploading assignments to external sites. Please review the examples of plagiarism provided by the BU Computer Science department. All suspected cases of plagiarism will be reported to the Academic Dean.

Accommodations

Boston University is committed to providing reasonable accommodations to students with documented disabilities. Students seeking accommodations should contact Disability & Access Services (25 Buick Street, Suite 300; 617-353-3658) as early as possible in the semester. A new Faculty Accommodation Letter (FAL) must be requested each semester; DAS will send this directly to instructors.

Religious Observance

Students observing a religious holiday during regularly scheduled class time are entitled to an excused absence. Please notify the instructor in advance to make arrangements for any missed work.

Recordings

Recording of lectures requires instructor permission. Students approved for recording as an accommodation must limit use to personal study and may not share recordings.