anthropocene
learning path
◆ from floating point to differentiable simulation

Computational Physics

Physics on a machine that cannot represent most numbers, cannot take a limit, and has to finish. This path builds numerical solvers one at a time - Euler, Runge-Kutta, implicit, symplectic - and then keeps going into the paradigms that reframe the whole problem: spectral methods, variational integrators, Monte Carlo, and differentiable physics.

5 chapters14 lessons≈ 6.3 h
01

Numerical Reality

2 lessons · 38 min

What is a number, once it has to fit in 64 bits?

Before any solver, the substrate. Floating point is not "real numbers with a small error" - it is a different arithmetic, with its own failure modes. The distinction between a badly conditioned problem and an unstable algorithm is the single most useful idea in the field, and it belongs first.

02

Derivatives Without Calculus

2 lessons · 50 min

How do you differentiate a function you can only sample?

Finite differences are the bridge from continuous mathematics to something a machine can execute, and they carry the tension that runs through the whole path: the same h that reduces truncation error amplifies roundoff. This is where order of accuracy stops being notation and starts being a slope you can measure.

03

ODE Solvers, One by One

4 lessons · 110 min

Given the slope everywhere, how do you find the path?

Every integrator here is a different answer to one question: you know f, and you want y. We build them in order - Euler, Heun, RK4, then implicit methods - and each one exists because the previous one failed in a specific, reproducible way you will have seen for yourself.

Why does a worse method give a better orbit?

RK4 is fourth order; velocity Verlet is second. Over a long orbit Verlet wins anyway, and not by a little. This chapter is where accuracy stops being the only axis - some methods approximate the trajectory, and others preserve the geometry the trajectory lives in.

05

The Other Paradigms

4 lessons · 120 min

What if stepping through time is the wrong idea entirely?

Everything so far marched forward in small steps. These do not. Spectral methods solve in frequency space; variational integrators discretise the action instead of the equations; Monte Carlo abandons determinism to beat dimensionality; differentiable simulation makes the solver itself something you can optimise through.