COREMat Framework

COREMat (COmputable REgression for Materials) is a Python framework for automated machine learning on tabular materials-science datasets — alloy research, materials-property prediction, and similar regression workloads. It automates the full workflow behind a single configuration: data ingestion, preprocessing, nested-CV training with hyperparameter optimisation, iterative feature selection, SHAP-based explainability, OOF/Hold-out evaluation, and interactive HTML reporting. The model registry ships 65 algorithm variants (gradient boosting, kernel methods, linear models, neural nets, foundation models) behind a unified ModelSpec interface.

What You Get

Every trained model gets a self-contained, interactive HTML report — open it in a browser, no server or notebook required. Excerpt of what's inside:

Every report is built post-hoc from persisted per-round predictions — re-opening a report never re-runs a model.

Installation

Requirements: Python >= 3.12, < 3.13

From PyPI (recommended — available from the first public release):

pip install coremat
# or with uv:
uv pip install coremat

From the Uni Rostock GitLab Registry (alternative, requires GitLab account):

uv pip install git+https://gitlab.uni-rostock.de/cs1302/coremat.git@main

Developers (editable install):

git clone https://gitlab.uni-rostock.de/cs1302/coremat.git
cd coremat
uv venv && uv sync
source .venv/bin/activate    # Linux/macOS
.venv\Scripts\activate       # Windows

TabPFN Authentication

TabPFN models require a free PriorLabs account. Set up once after installation:

coremat auth tabpfn

Getting Started

The documentation (mkdocs) is the single source of truth for usage — it covers the full quick-start walkthrough, sample data generation, and the resulting artifact layout. Start at docs/index.md or the Minimal Run guide.

Documentation

Building the docs

uv add mkdocs mkdocs-material
uv run -m mkdocs serve                          # local preview at http://127.0.0.1:8000, live-reloads on edit
uv run -m mkdocs build --strict --site-dir /tmp/site  # artifact build; --strict turns broken links/nav into errors

The pages CI job runs the same --strict build as a documentation lint — GitLab Pages is not hosted on this instance, so nothing is deployed automatically. Publishing the built site/ output anywhere remains a manual step until a hosting target is chosen (see the CI/CD section in docs/development.md).

Citing COREMat

If you use COREMat in your research, please cite it. Machine-readable citation metadata is provided in CITATION.cff at the root of this repository.

Acknowledgements

This project was developed using an agentic AI-assisted workflow with Claude Code by Anthropic.