Page 79 - ESAM-1-1

Engineering Science in
Additive Manufacturing ML in additive manufacturing

ML models more realistic and generalizable to real-time minimizing a loss function using an optimizer. When the
monitoring and control applications without requiring data quantity is small relative to the number of weights,
large and diverse datasets. 111,113,114 Compared to data-driven the system becomes highly underspecified, resulting in
methods that rely entirely on data to extract knowledge, numerous possible weight combinations that can closely
physics-informed learning incorporates the governing fit the limited training data. However, only a few of these
physical principles of the system, making it inherently combinations generalize well to the entire population.
more data-efficient. Consequently, highly complex models are prone to
overfitting the training data.
The most common approach to integrate physical laws
governing the complex AM phenomenon into the learning In AI-driven AM research, the growth in training data
process involves integrating customized loss functions and has lagged behind the increasing complexity of ML models,
constraints to effectively guide the training process. In as reflected by the growing number of weights. Collecting
129
this regard, FFNNs and their variants are usually modified and annotating AM in-process data remains constrained by
107
for the learning process. Another approach involves the high costs of experiments and measurements, limiting
directly incorporating physical equations into the loss dataset sizes. This widening gap between data availability
function or using them to derive physically informed and model complexity hinders the generalizability and
features as ML inputs. Similarly, high-fidelity simulations reliability of AI-driven AM systems. While acquiring large
109
driven by physics-based models are used to generate datasets remains challenging, methods such as feature
data for the training process. This includes using finite engineering, data augmentation, active learning, and
130
element analysis simulations or other computational knowledge transfer have emerged over the past 8 years to
physics models to generate large datasets that reflect partially bridge this gap. Figure 9 illustrates the methods
realistic scenarios and variability of the manufacturing that expand and mitigate this gap.
processes. Advanced approaches to inform the learning The gap between available data and model complexity
126
process through physics have focused on integrating in AI-driven AM has expanded primarily due to
multi-scale physical phenomena or combining several increased input dimensionality and more intricate model
physical domains (thermal, mechanical, and material) into architectures. High input dimensionality often necessitates
a single empirical model. Finally, knowledge descriptors larger neural networks with more trainable weights. Since
and process understanding have been used to design 2016, the introduction of image and time-series data as
and develop ML architectures and pipelines in AM input in AI-driven AM has increased dimensionality
applications. 130 compared to traditional tabular data. In 2017, point cloud
12
data – featuring even higher dimensionality – was utilized
6. Challenges and opportunities in research for defect detection. Multi-scale and multi-modal learning
and development methods further compound this issue by integrating data
of varying scales and modalities at the input layer.
Recent trends in applying ML to address existing
challenges in AM provide insights into future research As mentioned earlier, the timeline of AI-driven AM
opportunities. This section highlights how bridging the model adoption indicates a trend toward increasing model
gap between data and models can help realize the potential complexity and capacity. Early implementations, such as
of ML. Through the development of large AM datasets, ensemble methods, CNNs, and RNNs in 2018, employed
AM-specific foundation models, and efficient and scalable relatively moderate numbers of parameters suited for
ML model development, the next phase of ML applications smaller datasets and simpler tasks. Afterward, increasingly
can effectively harness the synergy between advanced complex architectures have been adopted (RCNNs, GANs).
computational techniques and domain-specific knowledge Most recent applications include complex architectures,
which will drive innovations across the entire AM process such as transformers, diffusion models, and LLMs, which
chain. were adopted in 2023, followed by VLMs in 2024. These
advanced models, known for their massive parameter
6.1. Bridging data and models counts and high capacity, enable sophisticated feature

With the adoption of advanced ML algorithms in AI-driven extraction and multimodal data integration. However,
AM, the increasing capacity and complexity of models this also underscores the growing need for large-scale,
drive higher demands for data quality and quantity. High high-quality datasets to mitigate the risks associated with
model capacity and complexity lead to a large number overfitting and to enhance generalizability.
of weights that define the mapping between input and Methods to address the gap between data availability
output variables. Typically, these weights are optimized by and model complexity in AI-driven AM focus on

Volume 1 Issue 1 (2025) 12 doi: 10.36922/ESAM025040004