Physics-Informed Classical and Quantum Neural Networks for One-Dimensional Schrodinger Eigenvalue Problems

At the heart of how we understand the universe lies a single, stubborn challenge: predicting how tiny particles like electrons move and settle around atoms. In the world of the very small, particles do not follow the smooth, predictable paths of planets or baseballs. Instead, they exist as waves of probability, described by a mathematical rule known as the Schrödinger equation. This rule tells us exactly where a particle is likely to be found and how much energy it holds, but solving it is notoriously difficult. For simple, idealized situations, scientists have known the answers for nearly a century. But for the complex, messy systems that make up real matter, the equations become so tangled that even the most powerful supercomputers struggle to find a solution. This is where a new generation of tools is stepping in, borrowing ideas from artificial intelligence to untangle the physics of the subatomic world.

Researchers in Pakistan and Mexico have recently tested two of these new tools on three classic problems that serve as the training ground for any new solver. They wanted to see if a standard computer program and a specialized quantum computer program could both learn to solve the Schrödinger equation without being fed a single example of the correct answer. Instead of memorizing data, these programs were taught the laws of physics directly. They were given the rules of the game—the equation that governs the particle's behavior, the requirement that the particle must exist somewhere in space, and the rule that different energy states cannot overlap—and asked to figure out the rest. The goal was to see if these "physics-informed" networks could find the precise energy levels and wave shapes of a particle trapped in a box or moving in a smooth curve, matching the known answers with high precision.

The team focused on three specific scenarios that physicists use to test their methods. First was the harmonic oscillator, a model of a particle bouncing back and forth in a smooth, curved valley. This is a gentle problem where the particle's wave shape changes gradually. Next, they looked at the infinite square well, a box with walls so high that the particle cannot escape, forcing the wave to hit zero abruptly at the edges. Finally, they examined the finite square well, a box with walls of limited height, allowing the particle to leak slightly into the forbidden region outside. These three cases cover a wide range of difficulties: smooth curves, sharp corners, and the tricky business of particles escaping their confinement.

To solve these, the researchers built two different types of learning systems. The first was a classical neural network, a standard computer program designed to mimic the way biological neurons connect and fire. The second was a quantum neural network, which runs on a simulated quantum computer. This second system uses the strange properties of quantum mechanics, such as superposition and entanglement, to explore a vast landscape of possibilities simultaneously. Both systems were trained using a composite score that penalized them for breaking any of the physical rules. If the wave function did not satisfy the Schrödinger equation, the score went up. If the wave did not fit inside the box or if two different energy states looked too similar, the score went up. The systems adjusted their internal settings over thousands of attempts to lower this score, effectively teaching themselves the solution.

The results showed that both methods worked remarkably well. For the smooth harmonic oscillator, both the classical and quantum networks found the energy levels of the first four states with an error so small it amounts to a few parts per million. They reproduced the exact wave shapes almost perfectly, matching the known mathematical solutions. When the researchers moved to the square wells, where the walls are sharp and the physics is more abrupt, the networks still performed with high fidelity. They recovered the correct energy levels and wave functions even where the potential changed instantly, a task that often trips up traditional numerical methods that rely on dividing space into a rigid grid.

A key difference emerged when the researchers looked at the more difficult, higher-energy states. As the particles became more excited, moving into states with more complex wave patterns, the classical network began to struggle. It found it harder to navigate the complex landscape of possibilities to find the best solution, often getting stuck in local errors. The quantum network, however, handled these higher states more reliably. Because the quantum circuit can explore a continuous range of transformations across its entire system, it maintained its ability to find the correct solution even as the problem became more intricate. This suggests that while classical computers can solve simple quantum problems with great accuracy, quantum systems may hold a distinct advantage as the complexity of the physical system increases.

The study confirms that these physics-informed approaches are a viable way to solve quantum problems without needing massive amounts of pre-calculated data. By embedding the laws of physics directly into the learning process, the networks avoid the need for the rigid grids that traditional methods require, making them naturally suited for problems where space is difficult to divide neatly. While the current tests were limited to one-dimensional systems with known answers, the success of both the classical and quantum versions points toward a future where these tools can tackle the complex, multi-dimensional systems found in real materials and molecules. The quantum approach, in particular, showed a robustness that hints at a growing capability to handle the most challenging calculations in quantum physics, offering a new path forward for understanding the building blocks of nature.