Neural network predicts how titanium alloy grains round off during annealing

What this research was about

Titanium alloys owe much of their behaviour to the shape of the crystal regions inside them. In Ti-55511, one of these regions is called the alpha phase, and after hot working it often appears as thin plates, or lamellae. Heating the metal afterwards lets those plates break up and round off into more compact grains, a process known as globularization. How far it goes depends on at least three things at once: how much the metal was deformed beforehand, how hot the later heating is, and how long it lasts. Mapping that three-way dependence usually means making and examining a great many samples.

The researchers used a geometric trick to compress that effort. By rolling a wedge-shaped sheet, thicker at one end than the other, they imposed a smoothly varying amount of deformation along its length, from no strain up to a true strain of 1.10. A single sheet therefore carried many different prior strains. Strips cut from it were annealed at four temperatures between 750 and 825 °C for times from 10 to 480 minutes, and electron microscope images taken at known strain positions were analysed to measure how much of the alpha phase had rounded off.

Where AI came in

The artificial intelligence here is a small neural network of a long-established type: a backpropagation network, meaning one trained by feeding errors backwards to adjust its internal weights. It had two hidden layers with four hidden units, and its starting weights were chosen by a genetic algorithm, a search method that breeds and mutates candidate solutions. It took three numbers as input — prior strain, annealing temperature and annealing time — and returned one number, the fraction of alpha phase that had globularized. Eighty measured records were used for training and sixteen, those from the 30-minute anneals, were withheld for checking; on those the average absolute relative error was 3.22 per cent.

The network stood in for a fitted equation of the sort that would otherwise be used to describe such kinetics. Once trained, it was put to two uses that produced results no measurement supplied directly. Each input was nudged up and down by small percentages to see which mattered most, giving an order of annealing temperature first, then prior strain, then time. The network was then run across grids of strain and temperature to draw kinetic curves — globularization against time from 10 to 480 minutes — filling in combinations that were never annealed and examined in the laboratory. The underlying mechanisms, by contrast, were established from the microscope work, and the strain field itself came from a conventional physics simulation rather than a learned model.

Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.