On a sound mixing console, there are hundreds of volume knobs, and each one affects its neighbors. Turn one, and the whole picture changes, but you can only check the result by listening to the entire chord, which takes hours. Designing a laser pulse for inertial confinement fusion is much the same. A tiny capsule of frozen fuel is blasted with powerful lasers, compressing it to temperatures and densities where nuclei fuse. The pulse shape determines whether the compression is smooth or destructive, but searching for it blindly, trying options one by one, is far too expensive.
In the new work, physicists replaced the "black box" of an expensive code with a trainable model that not only computes the outcome but also shows how each turn of a virtual knob affects the final goal. It's like descending a mountain in fog: instead of stumbling around, you get a map of the slope and can step downhill right away. The model was run for a facility on the scale of OMEGA (energy 25 kJ), where the laser pulse was described by 500 points, and the target geometry varied slightly.
This approach, differentiable modeling, turns an exhausting search into a fast descent to the best solution. For now, the model is simplified: it only accounts for external pressure, without full hydrodynamics. But even in this form, it finds pulses that would previously have required months of computation. The next step is to add more physical details and teach the algorithm to handle noise in the data, so the method can be applied to real experiments. Then tuning a thermonuclear explosion will become as predictable as mixing sound on that very console where it all began.
A typical target for inertial fusion is a hollow sphere about 2 mm in diameter, with a layer of deuterium-tritium ice frozen inside to a thickness of about a human hair.
In Michael Crichton's novel "Sphere," an alien spacecraft uses a similar principle of plasma confinement, but it is controlled not by lasers but by fields of unknown nature.