Posts

Author

John Kitchin

Posts from 2009 to 2025 are in the archive, at their original addresses.

How we do agentic science

agents

scimax

optimization

Our approach to agentic science, told through our software: scimax for VS Code, knowhow, litdb, crucible and fairchem-mcp, and the agent-facing features of difflow, pounce, discopt and the discopt plugins.

Oct 29, 2026

14 min

Surrogate models you can optimize globally

python

pycse

optimization

surrogates

Piecewise-linear surrogates from pycse, jax-ldt, jax-difftree and jaxsr can be written exactly as mixed-integer programs, so a MIP solver can find their global optimum with a certificate.

Oct 28, 2026

24 min

How a neural network builds features

machine learning

python

A neuron makes a linear combination of its inputs and bends it. That is the whole vocabulary, and it has consequences for products, ratios, units and depth.

Oct 23, 2026

22 min

Interpretations of what a neural network is

machine learning

python

A Fourier series is a one-hidden-layer neural network, and every neural network is a learned nonlinear feature transform followed by a linear model.

Oct 21, 2026

17 min

A zoo of differentiable surrogates that know their uncertainty

machine learning

uncertainty quantification

jax

surrogates

Symbolic models, trees, neural networks and flows from jaxsr, jax-ibart, jax-ldt, jax-difftree and pycse.sklearn: what each is, how it gets its uncertainty, and where to find an example.

Oct 20, 2026

23 min

One differentiable model, many tasks

optimization

jax

discopt

difflow

digital twins

Simulation, estimation, reconciliation, design of experiments, optimization, control and planning are different questions asked of the same differentiable model f(x; θ).

Oct 18, 2026

2 min