Résumé : A model order reduction approach combining reduced basis (RB) projection and substructuring by domain decomposition is developed to tackle nonlinear elasto-plasticity in parametrized coupled thermo-hydro-mechanical (THM) systems. The region-specific occurrence of plasticity in the THM process is exploited in domain decomposition to facilitate the simplified construction of localized reduced subspaces in order to perform projection on a multi-domain basis. Traditional RB projection is leveraged in the elastic zone while the computationally expensive nonlinear iterative procedure is confined in the zone where plasticity is assumed to be restricted. The applicability of the approach to parametric problems is investigated through inverse identification of material parameters involving a nonlinear THM model system for nuclear waste repository applications. The model reduction strategy is observed to significantly reduce the computational effort involved with up to 12 times faster simulation compared to finite element (FE)-driven inverse identification procedure.