Page 37 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

bone modulus, single-objective inverse optimization
111
provides efficient solutions by identifying process
parameters that achieve those targets. The core of the ML
implementation path is to construct an inverse search
framework driven by the performance target. This involves
first training a high-accuracy forward surrogate model that
maps process parameters to the target performance. Then,
with the goal of making the model’s predicted value as
close as possible to the desired performance, optimization
algorithms are used to perform a global search within the
process parameter space, inversely deriving the optimal
combination of process parameters.
112
Narayana et al. explicitly aimed to determine the
optimal process parameters for achieving extremely high
density and a specific build height for Ti-6Al-4V in DED.
Figure 10. Illustration of Pareto front with 3-objective problem 103 By training an ANN as an accurate forward predictor and
coupling it with an optimization algorithm for inverse
sorting, and crowding distance calculation mechanism, search, they inversely solved for the optimal combination
this algorithm can stably find a Pareto front with good of laser power, scan speed, and other parameters that would
convergence and high diversity, comprehensively revealing meet these predefined geometric and density requirements.
113
the trade-off relationships between performances. It is Similarly, the work of Nguyen et al. focused on
one of the most widely used algorithms in AM; 104,105 (ii) identifying the process parameter set that would achieve
multi-objective PSO: known for its fast convergence speed, the target near-full density (>99.8%) in LB-PBF Ti-6Al-4V.
this algorithm guides the search direction by maintaining Using a deep learning model to inversely explore the
the personal best and global best positions. Although process parameter space, they determined the specific laser
sometimes slightly inferior to NSGA-II in terms of the power, scanning strategy, and other parameters that would
uniformity of the solution distribution, it is an efficient minimize pore formation while maintaining the desired
choice when computational resources are limited and rapid density threshold. In scenarios aiming to maximize specific
acquisition of approximate optimal process parameters mechanical properties (such as surface hardness and
is desired; (iii) multi-objective accelerated process tensile strength) of implants, ML can likewise circumvent
106
optimization (m-APO): an emerging efficient framework the limitations of traditional experience. Gan et al.
114
that significantly reduces the number of experiments employed a SOM to inversely identify the optimal process
required to find a satisfactory Pareto solution set by window corresponding to target microhardness values in
skillfully leveraging prior knowledge and an innovative DED nickel-based superalloys. Their approach mapped
space-filling search mechanism. It is particularly from desired hardness values back to the optimal powder
suitable for rapid process adaptation; and (iv) multi- feed rate and other process parameters that would achieve
107
objective Bayesian optimization (MOBO): especially those targets (Figure 11A and B). This work demonstrates
suitable for optimizing expensive-to-evaluate black-box the potential of unsupervised learning techniques such as
functions. 95,100,108 When the performance prediction model SOMs in identifying the process window for biomedical
itself is computationally expensive to train, or when metal AM through reverse identification.
the optimization process needs to be directly coupled Furthermore, melt pool stability directly determines
with high-fidelity simulations, frameworks combining
GPR with MOBO can inversely locate optimal process the microstructure and defect state of the fabricated part.
115
parameters with minimal evaluations by intelligently Tapia et al. developed a surrogate modeling framework
balancing exploration and exploitation, demonstrating based on GPR that could inversely guide the selection
high sample efficiency. 109,110 of LB-PBF process parameters of 316L stainless steel.
Specifically, to achieve a stable and appropriately sized
3.3. Single-objective inverse optimization melt pool, the model could output the specific laser power
and scanning speed combinations that would produce the
When the design of biomedical metals has an extreme
requirement for a single core performance indicator, such desired thermal characteristics.
as load-bearing sites requiring extremely high fatigue life, Single-objective inverse optimization still holds an
or the need for precise control of elastic modulus to match important position in the AM of biomedical metals,

Volume 1 Issue 4 (2025) 15 doi: 10.36922/ESAM025440031