Page 44 - ESAM-1-4

Engineering Science in
Additive Manufacturing Machine learning for biomedical metal AM

controlled a laser metal deposition process by identifying a In forward prediction, ML functions as an efficient
linear state-space model. Building on a simplified physical surrogate model, accurately constructing nonlinear
model of the melting pool, Cao et al. proposed a control- mappings from process parameters to macrostructure,
154
oriented multi-input multi-output model and subsequently microstructure, and mechanical properties. This lays the
designed a tube-based robust multivariable MPC. In foundation for achieving first-time-right manufacturing.
recent years, data-driven nonlinear MPC has emerged as For inverse optimization, ML has driven a paradigm
a cutting-edge approach, significantly enhancing modeling shift from trial-and-error experimentation to intelligent
and control capabilities for complex dynamics. decision-making. Through sophisticated algorithms, it
In the field of data-driven MPC, Chen et al. proposed enables efficient searches and multi-objective trade-offs
155
a digital twin framework based on time-series DNN can be achieved within high-dimensional parameter
and MPC. The study employed the time-series dense spaces, thereby determining optimal process solutions
encoder as a surrogate model, enabling precise prediction for biomedical metals that meet complex clinical
of the entire future time domain for both molten pool requirements. In quality control, ML has achieved a leap
temperature and depth. On this basis, MPC dynamically from passive detection to active intervention by integrating
adjusts laser power to achieve precise temperature tracking diverse sensor data. Through real-time identification and
while strictly constraining molten pool depth within the diagnosis of process anomalies, it establishes closed-loop
ideal dilution range of 10–30%. This approach has been adaptive manufacturing capabilities, thereby paving a
shown to proactively suppress pore defect formation at its viable technical pathway for zero-defect production of
source. In comparison with conventional proportional- biomedical metals.
integral-derivative control methodologies, this framework 5.2. Current challenges
not only demonstrates equivalent temperature tracking
capability (R > 0.99) but also generates smoother, less Despite significant progress in ML for AM biomedical
2
fluctuating laser power curves. It inherently handles multi- metals, its further development and application still face
variable constraints, highlighting the immense potential critical bottlenecks:
of data-driven models in closed-loop control for complex (i) Dual challenges of data quality and quantity: The lack
manufacturing processes. The capability of the model to of high-quality, large-scale datasets remains a primary
handle multivariate constraints further demonstrates the factor constraining model performance. Experimental
application potential of data-driven methods in the field of data on biomedical metals is extremely costly to
closed-loop control for complex manufacturing processes. acquire, and inconsistent data standards across
research institutions create severe data segregation.
For the manufacturing of demanding biomedical
metals, it is essential to ensure that the microstructures (ii) Insufficient model generalization: Although existing
models typically perform well under specific
and macroscopic properties of the produced materials material and process conditions, their generalization
comply with stringent biomedical requirements, thereby
maximizing the service life. Through ML-enhanced capabilities decline significantly when confronted
closed-loop control and adaptive manufacturing, the with new material systems or process changes. This
restricts the adoption of ML solutions in widespread
qualification rate and performance consistency of ML industrial scenarios.
biomedical metals products can be significantly improved, (iii) Balancing physical consistency and interpretability:
while reducing reliance on operator expertise. Ultimately, Most current ML models remain black boxes, with
this speeds up the process of bringing highly reliable,
customized biomedical metals to market, paving the way weak correlations between their predictions and
for a more autonomous and intelligent future. underlying physical mechanisms.
(iv) Complexity of multiscale modeling: Cross-scale
5. Summary and outlooks modeling spanning from microstructural evolution
to macroscopic properties remains a major challenge.
5.1. Summary Integrating physical information across different
This review examines and elaborates on the role and scales and establishing accurate mapping relationships
advancements of ML in the field of AM for biomedical requires in-depth research.
metals, particularly in three core areas: forward prediction, (v) Technical barriers for real-time applications:
inverse optimization, and quality control. ML is emerging Deploying ML models for online quality control and
as a key driver for understanding and optimizing this real-time process adjustments faces multiple technical
complex manufacturing process, propelling the technology hurdles, including computational efficiency, latency
from experience-dependent to data-driven processes. requirements, and system integration.

Volume 1 Issue 4 (2025) 22 doi: 10.36922/ESAM025440031