Page 36 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

LB-PBF, instead of using process parameters directly, they to becoming trapped in local optima, failing to discover the
utilized a validated 3D heat transfer and fluid dynamics global optimal process window. This means potentially
98
model to compute six feature variables closely related to the missing parameter combinations that could achieve higher
balling formation mechanism, such as volumetric energy performance or better overall benefits.
density, surface tension, and melt pool aspect ratio—as Against this backdrop, optimization frameworks driven
inputs for ML. Using a genetic algorithm, they constructed by ML, by integrating performance prediction models
a balling susceptibility index prediction model, which with intelligent search algorithms, 99,100 enable precise and
achieved approximately 90% prediction accuracy tested efficient mapping from the performance space back to the
against 166 experimental datasets, with its effectiveness process parameter space, ultimately facilitating intelligent
further validated through LB-PBF experiments.
decision-making from performance requirements to
In defect prediction, the usage of ML is evolving process recipes.
from post-detection toward pre-hoc warning and virtual
twins. By generating virtual microstructures consistent 3.2. Introduction to optimization algorithms
with real defect statistics and leveraging explainable Within the inverse optimization framework, optimization
models to clarify causal relationships between process algorithms search the process parameter space to identify
parameters and defect types, ML paves a new path toward combinations that achieve target performance metrics.
achieving comprehensive, forward-looking quality control Different algorithms, with their unique mechanisms,
throughout the entire biomedical metals AM process. are suitable for different application scenarios. 101,102 For
inverse optimizing a single key performance objective,
3. ML-driven inverse optimization of AM algorithms strive to find the global optimum within the
process parameter space. Evolutionary algorithms such as genetic
3.1. Overview algorithm and particle swarm optimization (PSO), which
simulate natural selection or collective behavior, perform
After achieving accurate prediction of the process- global exploration in complex, non-linear spaces and can
structure-property relationships, the core challenge in effectively avoid local optima, making them common
AM process research shifts from an analytical problem choices for solving such problems.
to an inverse problem of greater engineering value. In
biomedical metallic AM, inverse optimization denotes However, optimization in AM is inherently a multi-
ML-based algorithms searching within high-dimensional objective optimization (MOO) problem. The core challenge
process parameter spaces to identify optimal parameter of MOO lies in handling these conflicting objectives to
combinations or Pareto fronts. Its core output comprises find the best compromise solutions. The goal is no longer
either a set of process parameters meeting target to obtain a single optimal solution, but to identify a set of
performance criteria or a process window achieving multi- non-dominated solutions known as the Pareto optimal
objective trade-offs, thereby bridging material design with set. These solutions collectively form the Pareto front,
applications. the set of all Pareto optimal points in the objective space,
which clearly delineates the trade-off boundaries between
Traditional optimization pathways face three major
challenges: (i) high-dimensional complexity: the AM different performance metrics beyond which no further
improvement is possible without worsening another. Each
process involves numerous parameters, such as laser solution on this front represents a specific performance
power, scanning speed, hatch spacing, layer thickness, and trade-off scheme; no single objective can be further
scanning strategy. These parameters interact in complex, improved without degrading at least one other objective.
non-linear, and coupled ways. This high-dimensional
97
parameter space makes optimization through traditional This allows decision-makers to select the most suitable
empirical methods or exhaustive search extremely process parameter combination from this set based on
difficult; (ii) cost and time constraints: full factorial specific clinical application requirements. As illustrated
103
design of experiments is a common method in traditional in the schematic by Hua et al. (Figure 10), within a 3D
objective space, both the search directions defined by
optimization, but the number of required experiments uniform reference vectors (hollow circles in the figure) and
grows exponentially with the number of parameters. For the actual Pareto optimal solution set (solid points in the
specific metal AM alloys or systems, budget constraints figure) are simultaneously presented.
make data-intensive methods difficult to generalize; and (iii)
susceptibility to local optimum: traditional optimization Commonly used MOO algorithms include the
methods based on single-point responses, such as one- following: (i) non-dominated sorting genetic algorithm II
factor-at-a-time or some local search algorithms, are prone (NSGA-II): noted for its elitist strategy, fast non-dominated

Volume 1 Issue 4 (2025) 14 doi: 10.36922/ESAM025440031