Enhancing AM part printability at the early design stage: a numerical and data-driven approach of DfAM

The overall aim of the research is to improve the manufacturability of additively manufactured (AM) part from the design stage. While the concept of Design for Additive Manufacturing (DfAM) has gained popularity over the past decade, the design principles are usually treated as empirical rules-of-thumbs and implemented manually. It results in difficulties in application to computational design and optimisation processes (e.g. topology optimisation, lattice generation etc.).

Furthermore, existing simulation and optimisation methods for AM often suffer from the formidably long computing time for large-scale or complex parts. This research tackles the problem by employing data-driven method (e.g. machine learning (ML)) to: 1) identify and incorporate the DfAM principles (as generalised from experimental data) into the numerical design optimisation processes; 2) achieve quick and accurate manufacturability simulation for large-scale parts at the conceptual design stage, allowing the parts to be adapted for AM early in the design cycle.

One main endeavour for realising the goals of the project is shifting the machine learning methods away from the traditional 'black-box' nature. The application of the 'physics-based machine learning' (PBML) concept is identified as a viable means for the objective. Inspired by the success of PBML in the field of aerodynamics, we propose embedding the governing equations during the AM process (e.g. heat dissipation equations to capture the thermomechanical behaviour during printing) in the machine learning model at the same time of training with experimental data to achieve fast yet physically accurate predictions of the outcome of printing, allowing subsequent design modifications/optimisations to be conducted accordingly.

The embedding of physical governing equations can be achieved through bespoke formulation of the loss function of the chosen ML techniques (e.g. deep neural network) and/or customised ML model architecture.

In addition to the development of physic-informed machine learning models, another focus is the integration of trained ML models with traditional numerical methods to leverage the strength of both techniques for multi-scale, high-fidelity, and rapid simulation as well as design optimisation. ML methods have superior speed at meso- to macro-scale since the computing time is de-coupled from the domain size, but depending on the training data used, they might be unable to produce accurate results at the microscale.

Selectively applying numerical methods at localised regions where greater details are needed can effectively resolve the issue without incurring too much extra computation effort as it is only run at the microscale.

The outcome of the research can benefit any AM-related applications by increasing the likelihood of 'first-time-right' and hence reducing the waste associated with failed prints and/or trial-and-error. Moreover, being able to generalise and implement DfAM rules in automated numerical design process reduces the reliance on experience to ensure successful AM, encouraging the adoption of AM technology and decentralised, automated manufacturing.

It will significantly benefit industries that have remote sites but require complex, high precision parts (e.g. offshore wind farms). With DfAM principles being encouraged/enforced, more efficient parts that leverage the manufacturing freedom of AM will be designed. It can lead to further reduction of environmental impact (e.g associated emission of vehicle parts) over the product's whole lifecycle.