A Deep Dive into Machine Learning for Material Science.
When we introduced the main applications of artificial intelligence (AI) in material science, we emphasized the importance of high-quality and sufficient data. In this article, we’ll explore how much training data is needed and where it comes from, using our work on AI-based microstructure analysis as an example. Specifically, we’ll focus on semantic segmentation through machine learning (ML).
The Process of Training and Applying an ML Model
To train an ML model, we need two key components: training data and ground truth. Training data consists of input data (e.g., microstructure images) paired with their expected outputs (e.g., annotated grain boundaries). This pairing is created by human experts through a process called annotation.
For microstructure segmentation, the input data is a microscopic image, and the ground truth is a manually annotated image highlighting the relevant features (e.g., grain boundaries). Once the model is trained, it can be applied to new, unseen data—a process known as inference.
How Much Training Data is Enough?
In computer science and the media, it is often said that thousands or even millions of data points are needed to train an ML model. Don’t worry, our ML models for structure segmentation don’t need that much data!
There are several reasons why this also works with less data.
It is worth comparing our microscopic microstructure images with images of everyday objects that are often used to train ML models. In contrast to these everyday images, our microstructure images are “richer” in information. For example, an image often contains significantly more than 50 grains, whereas these everyday images often only show one object (e.g. car/dog/cat). We can also use data augmentation (e.g. rotation) to create artificial variations that make the data set appear larger to the algorithm.
Another advantage is transfer learning, in which knowledge from previous tasks is transferred to new ones. Pre-training with data sets such as ImageNet helps the ML algorithm to recognize basic features (e.g. edges, shapes) that are also relevant for structural analysis.
Robustness and Generalization in Industrial Applications
For ML models to be effective in industrial settings, they must be robust and capable of generalization.
- Generalization ensures the model performs well on unseen data.
- Robustness ensures consistent results despite minor variations in input data (e.g., differences in sample preparation or imaging conditions).
Understanding Variability in Training Data
Metallographic processes such as sample preparation and image acquisition lead to variability and can result in significant differences in the final appearance of the microstructure. A holistic approach is essential here: we need to understand the processes and know the variances, only then can we determine the right images and the required amount of training data. Ultimately, the training data must cover all possible variations or be simulated by data augmentation.
The required amount of training data depends on the complexity of the microstructure and the level of variability. For simple structures with low variability, 50 images may suffice. For complex structures or high variability, hundreds of images are needed.
The Role of Domain Expertise in Annotation
Creating the ground truth data requires domain expertise. While simple structures (e.g., well-contrasted grain boundaries) can be objectively annotated, complex structures (e.g., bainite types or graphite morphologies) or incompletely contrasted grain boundaries can lead to subjective interpretations.
To minimize bias and to make some complex microstructures assessable at all using AI, correlative microscopy can be used. This technique combines several methods of analysis to provide a more accurate and objective ground truth. Although this requires additional effort during training, it ensures that the ML model works reliably during inference.
Conclusion The Importance of Domain Expertise and Data Quality
In summary, the success of AI-based microstructure analysis hinges on two factors:
- Domain expertise to understand and account for process variations.
- High-quality training data to ensure robust and generalizable ML models.
What’s Next? Exploring Transfer Learning and Pretraining
In our upcoming blog on AI-based microstructure analysis, we’ll delve deeper into Transfer Learning and Pretraining strategies. Key questions we’ll address include:
- Can ML models work effectively without pretraining on general datasets like ImageNet?
- Are domain-specific pretraining datasets (e.g., metallographic images) more effective for microstructure analysis?
- What can we achieve with our proprietary MECS “Grain Boundary Pretraining” framework, inspired by ImageNet’s success?
Stay tuned to learn how these techniques can reduce annotation effort while improving model accuracy for complex material science applications!
Ready to Leverage AI for Your Material Science Challenges?
Contact us today to learn how our AI-based microstructure analysis solutions can transform your research and industrial processes.