Cooperative data-driven modeling

Machine learning permeated almost every scientific discipline (Shanmuganathan, 2016; Wuest et al., 2016; Schmidt et al., 2019), and Solid Mechanics is no exception (Bessa et al., 2017; Capuano and Rimoli, 2019; Thakolkaran et al., 2022). With all their merits and flaws (Karniadakis et al., 2021), these algorithms provide a means to understand large datasets, finding patterns and modeling behavior where analytical solutions are challenging to obtain or not accurate enough. Focusing on plasticity modeling, its path-dependency posed a specific machine learning challenge that was recently addressed using recurrent neural networks, where time (or pseudo-time) can be naturally incorporated (Mozaffar et al., 2019). Since then several research groups proposed new neural network architectures and solved increasingly complex plasticity problems (Zhang and Mohr, 2020; Abueidda et al., 2021; Saidi et al., 2022; Bonatti et al., 2022). A similar trend is ongoing in other fields within and outside Mechanics (Liu et al., 2019a; Peng et al., 2021; Khalil et al., 2017; Dütting et al., 2019). Simultaneously, the scientific community is experiencing strong incentives to adhere to open science, with vehement support from funding agencies throughout the World to share data and models according to FAIR principles (Findable, Accessible, Interoperable and Reusable) (Wilkinson et al., 2016; Draxl and Scheffler, 2018; Jacobsen et al., 2020). There is also a clear need for end users to reuse these models and data. Nevertheless, there is a serious issue that obstructs the synergistic use of machine learning models by the community. Artificial neural networks, unlike biological neural networks, suffer from catastrophic forgetting (McCloskey and Cohen, 1989; French, 1999; Goodfellow et al., 2013).