Predicting the yield of a chemical reaction is a bit like trying to guess the harvest knowing only the soil composition. Traditionally, chemists have relied on long columns of numbers, quantum descriptors, that describe electron energies but say nothing about molecular shape. Yet sometimes two substances with nearly identical numbers behave differently simply because one is bulky, like a bouquet, and the other is flat, like a sheet of paper.
The authors gave the model eyes: alongside tabular data, they fed it simple 2D molecular diagrams — the kind you see in textbooks. The result was surprising: even a plain convolutional neural network trained only on these images predicts yields more accurately than the best methods using only quantum calculations. And when both information streams were combined via a cross-attention mechanism, the error dropped to 5.27%—a level previously thought unattainable.
Interestingly, the network architecture is designed to protect important numerical parameters from distortion when merged with images. For this, residual connections are used: they act like a safety rope, preventing electronic characteristics from getting lost in the visual data stream. The result is not just an accurate tool but also a transparent one: you can trace which parts of the molecule the network focuses on when making a decision.
The 2D molecular diagrams used in this work are the same "skeletal formulas" that chemists scribble on whiteboards. It turns out that even such a simple drawing is enough for a neural network to grasp spatial hindrances.