TUHH · Institute for Artificial Intelligence and Simulation in Mechanics
Machine Learning for Physical Systems: demos
Interactive demos for the lecture, winter semester 2026/27. Each runs in the browser, also on a phone, and needs no installation.
- Fit or overfit?
Fit a straight line by hand, then find out how flexible a curve should be to predict new measurements. - Watch a network learn
One bent line, then many: watch a small neural network learn how rubber stiffens, forward and back. - Be the tree
Pick the question that sorts eight concrete mixes best, grow a tree on 425 real mixes, then plant a forest and see what happens where the data end. - Where to test next?
Plan five concrete tests where the band is widest, drill five boreholes to find the gold, then see what the band of a Gaussian process assumes. - Physics in the loss
Put the law of physics into the loss: a network learns how a hot part cools and how a mass swings, then finds the cooling constant and the damping from a few readings. - Fields and rubber
Compare a classical solver with trained networks on a heated rod, see how a network learns ten periods of a swinging mass, then fit a rubber model with physics built in and check what it guarantees. - Stencils and images
Slide a kernel over a microstructure, watch one convolution learn the stencil of the heat equation, see how far one pixel sees, then let a small network read the conductivity of a tile and find out what it knows about shifts and rotations. - Diffusion in two dimensions
Add noise until three clusters of points become one round cloud, generate new points from pure noise one small step at a time, steer them towards one cluster, then check tiles drawn by the lecture's network with the solver. - Agent trace lab
Write the tool calls of a language model yourself and watch our program check and run them for a steel cantilever, then replay recorded runs of a small model step by step and see in twenty runs what input checks and a verifier change. - World model lab
Start a pendulum anywhere and watch a learned world model swing next to the true one, inside and outside the recorded swings, with the energy as a physical check, then see a predictor of camera frames turn a fly into a smear.
Historical recordings
- NETtalk learns to read aloud (Sejnowski and Rosenberg, 1986), audio on T. Sejnowski's page
- The Navlab vans of Carnegie Mellon University, with ALVINN steering (from 2:40), CMU Robotics Institute
- LeNet reads handwritten digits from a camera (1989), Y. LeCun's restored video
- Decision trees, interactive (MLU-Explain, J. Wilber and L. Santamaria)
- Random forests, interactive (MLU-Explain, J. Wilber)
- Transformer Explainer: GPT-2 small in your browser (A. Cho et al., Georgia Tech, 2024)
- Tiktokenizer: see how GPT models cut your text into tokens