Mechanica Labs
EN
Mechanica Labs

Solving complex problems transparently

Understanding the mechanisms underlying complex systems.

Fig. 1

We study how complex systems work on the inside. Our main work is mechanistic interpretability: we take trained neural networks apart to find the algorithms they learn. Knowing the algorithm lets us shrink a model, explain its decisions to the people who act on them, and gate what it is allowed to do. We also work on deep mines and mineral identification, where people need to understand a model before they act on it.

Research Publications

§ 1

Research

All research
  1. i Mechanistic interpretability We work out which algorithms neural networks learn, and test each explanation against the network. Knowing the algorithm lets us shrink a model, explain it to the people who use it, and gate what it can do.
  2. ii Deep mine systems An orebody is only partly known before it is mined. We follow that uncertainty through extraction, processing, tailings and backfill to find out which operating decisions it changes.
  3. iii Crystallography and mineral sensing Every mineral is a crystal whose atoms repeat in a lattice. X‑ray diffraction reads that lattice, and a microscope shows each grain in a rock. We build machine learning that identifies minerals from both, and we check how it reaches its answers.
§ 2

Approach

Inside a system that looks chaotic, some parts still follow exact rules. We look for those parts and write the rules down.

y py
Fig. 2 Poincaré section of the same system at E = 1/8. Each closed curve is a single orbit held on an invariant torus. The scattered points belong to chaotic orbits. Computed in your browser.

This plate comes from the Hénon–Heiles system, the model of a star moving through a galaxy that our logo is drawn from. Start an orbit almost anywhere and it scatters across the plate. Start it in the right place and it stays on a closed curve forever.

We look for the same kind of order in other systems. In a trained network it is the algorithm the network has learned. In a mine it is the part of the geological uncertainty that survives all the way to production.

About

§ 3

Selected publications

All publications
  1. Mining
    Geostatistical Discrete Rate Simulation of Coupled Mine, Processing, Tailings, and Backfill Systems

    Arthur Ayestas Hilgert, Alessandro Navarra

    42nd International Symposium on the Application of Computers and Operations Research in the Mineral Industry (APCOM 2027), Montréal Accepted

    Interactive figure

  2. Interpretability
    Deep neural networks divide and conquer dihedral multiplication

    Sihui Wei, Gavin McCracken, Gabriela Moisescu-Pareja, Harley Wiltzer, Doina Precup, Irina Rish, Jonathan Love

    International Conference on Machine Learning (ICML 2026)

    Interactive figure

  3. Interpretability
    On the Geometry and Topology of Representations: The Manifolds of Modular Addition

    Gabriela Moisescu-Pareja, Gavin McCracken, Harley Wiltzer, Vincent Létourneau, Colin Daniels, Doina Precup, Jonathan Love

    International Conference on Learning Representations (ICLR 2026)

    Interactive figure

  4. Interpretability
    Uncovering a Universal Abstract Algorithm for Modular Addition in Neural Networks

    Gavin McCracken, Gabriela Moisescu-Pareja, Vincent Létourneau, Doina Precup, Jonathan Love

    Neural Information Processing Systems (NeurIPS 2025)

    Interactive figure

§ 4

Our Team

Meet the team

  • Arthur Ayestas Hilgert

    Co-founder

  • Gavin McCracken

    Co-founder

  • Gabriela Moisescu-Pareja

    Co-founder

  • Sihui Wei

    Researcher