ML in Design & Process
Scientific ML
Multi-layer temperature history prediction during directed energy deposition using physics-informed neural network (PINN)
This project implements a physics-informed neural network (PINN)-based solution framework that predicts the thermal history during a multi-layer Directed Energy Deposition (DED) process.
- New opportunities for modelling the thermally induced distortion in metal AM with the meshless nature and the readily available derivative information from PINN solution.
- Overcomes the usual shortfall of neural networks (NNs) in dealing with discontinuities making multi-layer PINN-based simulation possible.
- Benchmark against ANSYS validates accuracy of the proposed framework to be comparable to numerical methods. Additionally, offering computational time-savings thereby making it amenable for use in design-optimisation algorithms.
The proposed framework sets the foundation for the subsequent exploration of applying scientific machine learning (SciML) techniques to real-life engineering applications. Furthermore, remarks on strategies to improve ease of training and prediction accuracy by PINN for the particular use case in DED temperature history prediction have been made.
▼ Relevant Publication(s)
Extending multi-layer temperature history prediction to more complex DED configurations using eXtended PINNs (XPINNs)
This work implements an eXtended PINNs (XPINNs)-based framework that utilises domain decomposition and demonstrates performance and capacity improvement over the PINN-based counterpart in multi-layer thermal history prediction of Directed Energy Deposition (DED) process.
- Significant improvement of prediction accuracy for multi-layer DED parts of simple geometry.
- Capability to account for complications such as dwell time and voids in design.
- Shares SciML-based framework’s meshless nature and immediate availability of derivative information.
- The scalable implementation of the framework allows a user defined number of sub-domains to be define. The code will be available shortly.
The proposed framework brings SciML-based framework for real-life DED applications one step closer with the expanded capability. Remarks on the appropriate level of domain decomposition are also made.
▼ Relevant Publication(s)
Data Driven Design
TO and ML-based optimisation to design FGMM
We propose a novel lattice generation method that designs Functionally Graded Matamaterials (FGMMs) with the assistance of Topology Optimisation (TO) and Machine Learning (ML).
- A Neural Network (NN)-based inverse lattice generator is trained to output lattice unit cells from the input of target mechanical properties.
- Utilises information from fast low-resolution Topology Optimisation (TO) design to inform the trained inverse generator and map lattice cells.
- The proposed method provides close to optimal performance whilst making the realisation of these solutions exceptionally quick.
Benchmark
We conducted a benchmark of our lattice generation approach by comparing with other approaches; dehomogenisation and inverse homogenisation.
- Our approach has comparable optimal structure’s performance and the computational efficiency, compared to dehomogenisation.
- Ensure manufacturability of the obtained structure without overhang angle or grey scale region.
- Versatile for multiple manufacturable unit cells, opening possibility in wider application fields.
ML-based inverse design of multi-material curved lattice structures
In this work, we proposed a ML-based inverse design framework to realise multi-material curved lattice structures, enabling the optimisation of multifunctional FGMs.
- The proposed unit cell design parameterisation enables simultaneous optimisation of shape, size and material, offering vast property space beyond conventional VF grading.
- High accuracy and efficiency in property prediction and inverse design by ML models, capable in handling mixed-type design features and properties across different physics.
- The MDN enables one-to-many inverse design, allowing further filtering based on manufacturability or other criteria.
Smooth transition unit cell generation via latent-space arithmetic
A latent-space-arithmetic-based framework is proposed and developed in this project to generate smooth transition between dissimilar unit cells. It allows lattice structures with multiple unit cell types to be realised. * Variational Autoencoder (VAE) is trained to map the unit cells to the latent space * The proposed strategy ensures both smoothness in transition and connectivity with the target cells. * Transitions between different types of triply periodic minimal surface (TPMS) and truss-based unit cells can be generated. Benchmark comparisons against analytical morphing and existing ML-based solutions indicate that the proposed framework consistently achieves superior performance.
Machine learning based lattice generation method derived from topology optimisation
We propose a novel lattice generation method that designs graded lattice structures with the assistance of Machine Learning (ML).
- A Neural Network (NN)-based inverse lattice generator is trained to output lattice unit cells from the input of target mechanical properties.
- Utilises information from fast low-resolution Topology Optimisation (TO) design to inform the trained inverse generator and map lattice cells.
- The proposed method provides close to optimal performance whilst making the realisation of these solutions exceptionally quick.
