Weather models: a ‘process-additive’ neural network uncovers errors that cancel each other out between clouds and convection

Each branch of the PANN network attributes a portion of the error to a distinct physical process, thereby revealing biases that cancel each other out and remain undetectable using conventional diagnostic methods.

In forecasting and climate models, two errors can cancel each other out: poorly represented convection can be masked by a cloud pattern biased in the opposite direction. The overall result seems correct, but for the wrong reasons, which makes it difficult to improve the models. Yuiko Ichikawa, Edward Groot and Jack Atkinson propose a tool to unravel these compensatory biases: the Process Additive Neural Network (PANN).

One branch per physical process

The principle: each branch of the network represents the trend error associated with a given physical process, and the sum of the branches must reconstruct the total error relative to a reference trend. Once the training is successful, each output indicates the contribution of a process to the error, and their average gives an estimate of the bias specific to that process. The first tests are case studies of the OpenIFS single-column model, which focus on clouds and convection, two processes so tightly coupled that their errors are difficult to separate.

The authors insist that the goal is not to replace human expertise on atmospheric physics, but to extend it to situations where the complexity of the interactions makes direct diagnosis impractical. This form of physics-structured interpretability should help design more coherent weather and climate models.