Raúl Larrubia Sobrino presents his Semester thesis on "Exploration and Benchmarking of MultiONet Architecture for Operator Learning"

Abstract:
Neural operators have emerged as a powerful framework for learning solution operators of parametric partial differential equations. Among these, MultiONet extends the DeepONet architecture by aggregating information across multiple network depths, improving predictive accuracy without increasing model capacity. In this work, several parameter-preserving strategies are investigated to further enhance the performance of MultiONet. Specifically, three directions are explored: agentic aggregation mechanisms that reorganize depth contributions, sequential training via dynamic loss scheduling, and modified training-data distributions with shared learning. Numerical experiments on the Darcy-flow benchmark show that none of the proposed approaches outperform the baseline MultiONet, but they provide valuable insight into the role of depth aggregation, optimization scheduling, and data exposure in operator learning. In particular, data distribution combined with shared learning yields modest accuracy improvements while reducing computational cost. These results clarify the limitations and potential of parameter-efficient extensions of MultiONet.