[2N4-GS-10x] AI application

1:30 PM - 1:45 PM JST(4:30 AM - 4:45 AM UTC)

[2N4-GS-10x-01] Pareto Physics-Informed Neural Networks for Constrained PDEs in Finance

〇Kentaro Hoshisashi¹ (1. University College London)

[[online]]

Keywords:

Derivatives pricing,Physics-Informed Neural Networks,Partial Derivatives Equations,Derivative-Constrained optimisation,Self-adaptive weighting loss balancing

We introduce a unified, constraint-aware, physics-informed framework for solving partial differential equations (PDEs) that (i) treats derivative-based inequality constraints on par with PDE residuals and boundary/initial conditions, (ii) casts training as \emph{multi-objective} optimisation with a Pareto-balanced descent, and (iii) specialises to implied volatility surface calibration in finance under no-arbitrage constraints. We also demonstrate the method on synthetic data for analysis of the training dynamics to Pareto-stationary points.