In 3D printing, performance fails quietly, as warped geometry, brittle layers, inconsistent strength, or a part that passes inspection but doesn’t survive real use. The reason is deceptively simple: in Fused Filament Fabrication (FFF), mechanical quality is shaped by thermal history. Every parameter choice (nozzle temperature, bed temperature, print speed) changes how heat flows, how polymer chains reorganize, and ultimately how layers bond.
That is the core challenge addressed by Vanaei, Moezzibadi, Tcharkhtchi and Khelladi in their Journal paper, “A hybrid genetic algorithm–neural network model for optimizing thermal and mechanical characteristics of 3D-printed parts” (DOI: 10.1108/RPJ-02-2025-0085). Instead of relying on trial-and-error or single-factor tuning, the authors build a hybrid optimization strategy that mirrors how modern manufacturing teams wish process engineering worked: learn from real production evidence, search efficiently for better configurations, and validate
that the “best settings” still make physical sense. In their approach, a neural network learns the relationship between printing conditions and outcomes, while a genetic algorithm explores combinations of parameters to find those that improve targeted mechanical properties. Crucially, this isn’t “AI replacing engineering judgment”. A heat-transfer model is used to check whether the optimized choices align with the physics of the deposition process, grounding predictions in first-principles reasoning.
This is where the paper aligns strongly with the AI-DAPT project. AI-DAPT’s ambition is to design and develop an AI-Ops framework for reliable, continuously improving data/AI pipelines, explicitly coupling science-guided models with data-driven learning, and keeping a human-in-the-loop for trust and accountability. The proposed paper exemplifies a hybrid modelling loop that combines data-driven learning (ANN surrogates), search-based optimization (GA), and physics-based validation (heat-transfer modelling) to constrain the solution space to physically plausible regimes. This is directly aligned with AI-DAPT’s goal of advancing hybrid science–AI
solutions supported by high-quality evidence, and of operationalizing them through lifecycle practices that improve robustness during deployment. In a manufacturing setting, such a loop can be embedded into an AI-Ops-oriented workflow where models are continuously updated as new production data arrive, deviations in predicted–observed properties trigger diagnostic checks, and parameter recommendations remain auditable through their linkage to first-principles constraints, enabling engineers to supervise, validate, and approve adaptations under a human-in-the-loop paradigm. Models should not only predict accurately, but also remain stable under changing operating conditions, be monitorable for drift, and be recalibratable when the data distribution shifts (e.g., new filament batches, different printer hardware, altered ambient conditions). The hybrid structure in the paper is a strong example of how such systems can be engineered so that updates are justified scientifically, supporting the type of continuous validation and human oversight that AI-DAPT aims to systematize across its industrial demonstrators.