<p>Physics-informed learning has been increasingly explored in geotechnical engineering, but its computational efficiency and practicability in engineering practice often incur scepticism. To address these limitations, this study enhances current physics-informed learning by integrating domain decomposition, transfer learning, data assimilation and customised optimiser and loss functions. Its feasibility is demonstrated by applying it to nonlinear time-dependent 1D and 2D Allen–Cahn equations, as representative phase-field model, including drop shrinkage, coarsening and kissing bubbles scenarios. The results indicate that time domain decomposition physics-informed learning enables solving the long time-dependent problems of interest, whilst using transfer learning can achieve a significant acceleration in computation and this effect increases with the increasing time domain. This framework allows for flexible data assimilation and its combination with transfer learning enables the utilisation of multi-source data, including sparse in situ monitoring and historical data, to derive accurate solutions for problems at hand. This capability offers a promising and cost-effective strategy for sensor installation, real-time prediction, model parameters identification and digital twinning in engineering practice.</p>