Simultaneous and Dynamic Prediction of Coupled Parameters in Colloidal Micronano Motion Systems by Physics-Informed Machine Learning

For probing micronano particle motion interface dynamics, intrinsic physical coupling among particle hydrodynamic radius (rh), interfacial viscosity (η), and temperature (T) hinders their simultaneous and dynamic acquisition and limits the labeled data available for conventional machine learning (ML). Here, we propose a physics-informed machine learning (PIML) paradigm that uses digital-twin simulations to endow ML models with physical priors. The digital twin reveals a millionfold disparity in parameter influence on motion signals and severe parameter unidentifiability in standard ML models. To overcome this, a hybrid ML architecture, RF-MBGA-Net, integrates a random forest (RF) with a multibranch gated attention (MBGA) network to balance parameter influence scales via intermediate prediction feedback and data augmentation. RF-MBGA-Net achieves high-accuracy predictions on simulated data (R2 > 99%; MAEs: rh= 3.9 nm, η = 0.023 mPa·s, T = 1.5 K). The trained RF-MBGA-Net performs simultaneous and dynamic prediction of the three parameters from experimental motion signals. The predicted rh, η, and T align with dynamic light scattering (DLS) results, rotational rheometer measurements, and sample-set temperatures. Moreover, the model captures inverse η-T coupling and molecular dissociation kinetics. The proposed PIML paradigm provides a scalable route for decoding coupled inaccessible parameters and probing micronano complex interface dynamics.