Skip to content

Case Studies

A case study is LeMat-Synth pointed at one scientific domain: the same pipeline, with domain-specific choices about which materials to look for, which plots are worth reading, and how results are written out.

Three are shipped with the repository. Each is a thin script — a few dozen lines — on top of two reusable pieces:

flowchart LR
    DC["DomainConfig<br/>(what this domain cares about)"] --> BR["BatchRunner<br/>(PDF discovery, retries, output)"]
    BR --> PIPE["SynthesisPerformancePipeline<br/>(the standard stages)"]

BatchRunner handles everything domain-independent — PDF and supplementary-file discovery, rate-limit-aware retries, resumable runs, progress reporting. DomainConfig supplies the four things that differ between domains. You never edit the runner; you assemble a config.

The three built-in domains

Domain What it extracts Script Guide
Thermocatalysis Synthesis + NH₃-decomposition conversion curves, benchmarked against human ground truth across several VLMs case_study_thermocatalysis/ Thermocatalysis
Superconductors Synthesis + critical temperature Tc, read both from text and geometrically from ρ(T)/R(T) plots case_study_superconductors/ Superconductors
Porous materials Synthesis + adsorption isotherms for MOFs, zeolites and COFs case_study_porosity/ Porous materials

Building a fourth one — electrochemistry, battery cycling, thermoelectrics, anything with a plot and a recipe — is covered in Building your own case study, and built end to end in Tutorial 7.

Running a built-in domain

Every case-study script takes the same two positional arguments and the same three flags:

uv run examples/scripts/case_study_porosity/run.py <pdf_dir> <output_dir> [flags]
Flag Effect
--max N Process only the first N papers — always do this first
--skip-existing Skip papers that already have results, so an interrupted run resumes
--skip-figures Text and synthesis only: no figure detection, no VLM, much faster and cheaper

Thermocatalysis is the exception — it adds a caching and evaluation harness on top, documented on its own page.

Important

The data/ directory is git-ignored, so no PDFs or ground-truth files ship with the repository. Every case study needs you to supply your own corpus. Tutorial 2 shows how to assemble one from the LeMat-Synth-Papers dataset.

Choosing a starting point

Use the built-in DomainConfig factory and point the script at your PDFs:

DomainConfig.for_catalysis()
DomainConfig.for_porosity()
DomainConfig.for_superconductivity(claude_model="claude-sonnet-4-20250514")

Start from that factory, then override the one piece that differs — usually the plot filter or the material prompt.

Work through Tutorial 7, which builds a thermoelectrics domain from an empty file and tests each piece as it goes, then keep Building your own case study open as the API reference.

Skip case studies altogether — lemat-synth batch over a folder of PDFs is enough. See the CLI Reference.