In the Data driven design course, I treated the design of a gearbox as an optimisation problem. Instead of choosing the gear ratios by hand, I wrote a Python program that describes a gearbox as a graph of shafts and gears, generates random gearboxes, and lets a genetic algorithm evolve them. Each gearbox has its own DNA, which is mutated and crossed over generation after generation to keep only the best designs.

To judge each gearbox, the program simulates a full acceleration from 0 to 200 km/h. It uses real engine curves (torque, power and fuel consumption), air resistance, rolling resistance, the efficiency of every loaded and unloaded gear and shaft, and the time needed to change gears. The fitness can be the time to reach top speed or the fuel consumed. Since thousands of simulations are needed, I ran them in parallel on every core of the processor. During the course, I also compared genetic algorithms, simulated annealing and Bayesian optimisation on other design problems.