Physics-informed multi-head LSTM-based predictive-prescriptive framework for process parameter optimization in fused deposition modeling

Additive manufacturing (AM) by fused deposition modeling (FDM) is prone to dimensional errors, warpage, and residual stress due to complex thermo-mechanical interactions. We present a predictive–prescriptive framework that combines thermomechanical simulations of filament-deposition physics with a multi-head Long Short-Term Memory (LSTM) network to predict layer-wise temperatures and quality metrics (deformation, strain, stress), and to refine process parameters that minimize defects. A 64-case factorial dataset of layer height, nozzle speed, and nozzle temperature was generated via coupled transient thermal–structural analyses for experimentation. The proposed framework produces (i) multi-probe temperature forecasts; (ii) global thermo-mechanical responses; and (iii) parameter prescriptions. It achieves probe-wise temperature RMSEs of 0.13–0.17 °C, deformation errors ≤ 0.02 mm, and stress errors within 0.3 MPa on the test cases. The outputs were validated through experiments involving re-printing guided by the framework, infrared thermography, and stress–strain testing. Using the refined process parameters, we observed a 21–26% reduction in residual stress during compression tests and a 11–39% decrease in total deformation. Meanwhile, infrared thermography showed temperature prediction errors of less than 3.5%. By connecting predictions with actionable parameter recommendations, this approach transforms quality assurance in fused deposition modeling from reactive inspection to proactive, physics-informed, and data-driven process parameter optimization.