Newsletter Topic #4 - AI-based surrogate models speed up the development of transparent wood-based materials

Artificial intelligence (AI) is a broad term referring to the capability of certain computational systems to perform tasks typically associated with human intelligence, such as reasoning, learning, creativity or pattern recognition. Recent progress in development of AI systems, particularly in their subset known as machine learning (ML), has been helpful in many fields of science and technology. The idea behind ML is that the computational system instead of being explicitly told what to do at each step, is trained with the available data to perform certain tasks it has been designed to. Basic structures used in ML are artificial neural networks (NNs), which consist of interconnected nodes called neurons. The neurons are usually organized in layers, starting from the input layer and ending with the output layer. The output value of each neuron in a given layer is determined by its input, which comes from the neurons in the preceding layer, and the internal parameters of the NN. The values of those parameters are adjusted during the training to minimize the difference between the NN’s prediction and the ground truth, which comes from the available refence data.

Fig. 1: Surrogating a physical model with a neural network.

AI-based models, once properly trained, can be used to perform certain tasks with a high accuracy and very often faster than other algorithms that are not based on AI. Such models that approximate the behaviour of the original model while being computationally cheaper are known as surrogate models (Fig. 1). Surrogate models can be useful in materials science in modelling certain structures and predicting their physical properties. This could in turn help to optimise desired properties with respect to the structural parameters, contributing to better design and improved performance of materials.In the context of transparent wood composites, AI can help in generating realistic wood microstructures and predicting their physical (mechanical, optical) properties. Modelling the complex wood microstructure comprising fibers, vessels and ray cells proceeds in several steps. The crucial step is distortion of the whole structure, which mimics the effect of the fiber growth on the neighbouring cells. However, this step is also very time-consuming, making generation of a whole set of various wood microstructures, which could be used for a further parametric study, impractical. This is an example of a task that can be accelerated by replacing the original model by a surrogate model based on AI. In a similar vein, replacing physics-based modelling of the wood composite stiffness parameters with AI-based surrogate can speed up the computations by orders of magnitude, enabling large-scale finite element analysis or optimization of manufacturing a wood composite for desired structural integrity. Finally, studying light interaction with wood composites can be carried out by integrating the underlying physics laws into an AI-based model, resulting in a significant speed-up compared to Monte Carlo simulations. To achieve all these benefits, three surrogate models have been developed.

Fig. 2: U-Net neural network for performing distortion of wood microstructures.

A surrogate model based on a U-Net NN has been developed to perform the distortion of wood microstructures (Fig. 2). The U-Net's architecture allows to perform image-to-image transformations, making it a good candidate for the task. After being trained on a sufficiently large dataset the U-Net is able to perform distortion of the wood microstructure in a shorter time than the original model [1]. This allows a fast generation of wood microstructures, which in turn could be useful in a parametric study of their physical properties, such as transparency or mechanical strength.

Fig. 3: Kolmogorov-Arnold neural network for predicting stiffness parameters of wood composites.

For computing structural strength parameters of wood composites, a surrogate model based on recently proposed Kolmogorov-Arnold networks combined with Bayesian statistics has been developed (Fig. 3). The surrogate can be up to millions of times faster in comparison to the original physics-based model [2]. The surrogate has been applied to optimization of microscopic parameters of wood composites where it has enabled the use of computationally intensive but parallel algorithms such as differential evolution.

Fig. 4: Illustration of photon transport in the Monte Carlo model.

Fig. 5: Schematic representation of the fully connected neural network architecture.

Building on these developments, our work takes AI-driven modeling into new territory by exploring how light behaves in transparent wood (Fig. 4). We've developed a physics-informed NN using data from Monte Carlo simulations, but here's what makes it different: it doesn't replace physics — it works with it [3]. The physics-informed ML-based model dramatically speeds up our calculations while still respecting the fundamental laws of how light interacts with materials. By combining simulated and experimental data, our model can accurately predict how much light passes through or reflects from transparent wood (Fig. 5), and it does this efficiently without losing touch with physical reality. For us, this represents what AI in science should be: a tool that accelerates discovery and helps us design more sustainable materials, while keeping our understanding firmly grounded in the actual principles of physics. In conclusion, AI-based models are promising tools for studying physical properties of novel materials with complex structures such as transparent wood composites. They offer significant speed-ups compared to more traditional approaches, while accurately reproducing the behaviour of the original models and respecting the physical laws. AI-based methods can be particularly useful in tuning the physical properties of studied systems, which can contribute to improved design and discovery of sustainable materials.

References

[1] M. Mińkowski, F. Seyedheydari, M. Seppi, B. Chen, S. Särkkä “Deep learning-based surrogate model for generation of transparent wood microstructures” (submitted)

[2] T. Härkönen, M. Königsberger, M. Mińkowski, J. Füssl, S. Särkkä “Partially-Bayesian Kolmogorov-Arnold network surrogate modeling of microscopic biocomposite stiffness” (submitted)

[3] F. Seyedheydari, M. Nasiri, M. Mińkowski, S. Särkkä “Determination of particle-size distributions from light-scattering measurements using constrained Gaussian process regression” (submitted)