Data assimilation is crucial for sensor and load characterization in fields such as Structural Health Monitoring. However, multistable stochastic systems pose challenges for traditional assimilation methods. We propose the use of a physics-informed machine learning strategy for identifying physically compelling stochastic nonlinear dynamics from data [1][4]. The algorithm is constrained by the Fokker-Planck equation to ensure the interpretability the solution. The formulation is complemented by the use of the framework of Gaussian Process Latent Force Models to formulate the load as a Gaussian Process [2]. We apply the method for the characterization of the response and properties of the multidimensional stochastic dynamics of bistable energy harvesters, characterized by nonlinear and mechano-electric coupled phenomena [3]. As a result, we obtain the reconstruction of the response surface of the mechanism regarding its parameters for exploration of combinations, and an updating algorithm for the hybrid twin of individual energy harvesters for continuous control and estimation of both drift and the stochastic load. REFERENCES [1] Callaham, J.L., Loiseau, J.C., Rigas, G. and Brunton, S.L.,. Nonlinear stochastic modelling with Langevin regression. Proceedings of the Royal Society A, 477(2250), p.20210092, 2021. [2] Vettori S, Di Lorenzo E, Peeters B, and Chatzi E. Assessment of alternative covariance functions for joint input-state estimation via Gaussian Process latent force models in structural dynamics. arXiv preprint arXiv:2306.16302. 2023. [3] Kumar, P., Narayanan, S., Adhikari, S. and Friswell, M.I.,. Fokker–Planck equation analysis of randomly excited nonlinear energy harvester. Journal of Sound and Vibration, 333(7), pp.2040-2053, 2014. [4] Chen, X., Yang, L., Duan, J. and Karniadakis, G.E.,. Solving Inverse Stochastic Problems from Discrete Particle Observations Using the Fokker–Planck Equation and Physics-Informed Neural Networks. SIAM Journal on Scientific Computing, 43(3), pp.B811-B830, 2021.