Abstract: We present a novel framework for predicting granular material properties, specifically shear strength and consolidation behavior, through automated multiscale convolutional analysis of X-ray microtomography µCT data. The method uses a hierarchical convolutional neural network HCNN architecture that extracts features directly from 3D µCT scans, eliminating manual feature engineering and enabling fast, high-throughput predictions. In comparative tests, the approach achieved an approximate 30 percent improvement in accuracy over traditional FEM methods, significantly reducing computational cost and processing time. Commercial applications include formulation and process optimization in construction, pharmaceutical excipients, and powder metallurgy.
Introduction: Granular materials such as sands, powders, and aggregates are critical in numerous industrial processes. Predicting macroscopic properties such as shear strength, consolidation, and permeability is essential for design and performance. Traditional methods such as FEM require extensive characterization and manual extraction of parameters such as particle size distribution, shape factor, and packing density. Advances in 3D µCT imaging offer detailed representations of the microstructure, but their efficient exploitation demands automated solutions based on artificial intelligence and custom software.
Background and related work: Previous studies have combined µCT analysis with FEM modeling, particle shape analysis, or statistical methods, often relying on manual feature selection. Convolutional neural networks CNN have demonstrated success in image recognition, but many applications have been limited to 2D or predefined feature extraction. This work develops a 3D HCNN designed to learn directly from µCT volumes and predict macroscopic properties without feature engineering.
Methodology: HCNN architecture: The core is a multiscale HCNN that extracts features at multiple levels. Preprocessing: µCT scans normalized to fixed resolution, particle segmentation, and noise cleaning. Initial layers perform voxel convolutions with 3x3x3 kernels to detect edges and local textures. Repeated convolutional blocks with batch normalization and ReLU activation extract higher-level patterns. Regional pooling reduces dimensions and increases the receptive field. Global average pooling transforms maps into representative vectors. Fully connected layers reduce dimensionality before the linear output layer that predicts shear strength in kPa and consolidation coefficient.
Mathematical details and training: The convolution operation is expressed as y = W * x + b applied over 3D cubes, followed by ReLU(x) = max(0,x). Max pooling 2x2x2 reduces resolution while retaining maximum response. Training minimizes MSE = sum((y_true - y_pred)^2)/n with the Adam optimizer and initial learning rate 0.001 with decay. Dropout regularization and early stopping were applied on the validation set to prevent overfitting.
Experimental design: Material: quartz sand with particle size distributions obtained by sieving. µCT scans at 2.5 µm resolution per voxel. Mechanical tests: triaxial compression according to ASTM standard and consolidation tests in an oedometer to obtain the consolidation coefficient. Dataset of 100 samples with diverse particles and packings, partitioned into 70 percent training, 15 percent validation, and 15 percent testing. MSE loss function and MAPE and R-squared evaluation metrics.
Results and discussion: The HCNN outperformed calibrated FEM simulations, achieving approximately 92 percent accuracy in shear strength (MAPE < 10 percent) and 88 percent in consolidation coefficient (MAPE < 12 percent), compared to 82 and 78 percent respectively for FEM. Computation time per sample was reduced from 15 minutes to 30 seconds using optimized GPU implementation. Analysis of convolutional filters indicates that the network learns to detect particle shape, packing density, and contact zones that correlate with macroscopic properties.
Technical interpretation: The key advantage is the automation of feature extraction from raw 3D data, enabling custom software pipelines and tailored applications for industrial processes. Limitations include dependence on µCT scan quality and greater generalization work for materials with radically different composition; data augmentation and active learning strategies are recommended.
Roadmap and scalability: Short term 1 year: expand dataset to pharmaceutical powders and metal powders, integrate with characterization platforms, and offer cloud API. Medium term 3 years: implement active learning and use of GANs to augment synthetic data. Long term 5 to 10 years: integrate into real-time control loops between µCT acquisition, HCNN prediction, and process control such as additive manufacturing, incorporating AI agents and reinforcement learning for autonomous optimization.
Practical applications and use cases: A cement manufacturer can virtualize screening of particle size distributions to optimize strength and workability. In pharmaceuticals, the method facilitates rapid selection of excipients and screening for tablet stability. In powder metallurgy, it allows adjusting particle size distribution for mechanical properties after sintering.
Verification and reliability: Validation through independent sets and statistical significance analysis comparing HCNN and FEM. Use of interpretability techniques such as activation maps to correlate regions of interest with physical variables. Optimized GPU implementations ensure parallel processing and reduced times, facilitating deployments in aws and azure cloud services.
Technical contribution: A 3D HCNN architecture designed for multiscale analysis is provided, capturing local interactions and global structure without the need for feature engineering. This enables the creation of business intelligence platforms and artificial intelligence solutions focused on materials, driving industrial digital transformation.
About Q2BSTUDIO: Q2BSTUDIO is a custom software and application development company specialized in artificial intelligence, cybersecurity, and aws and azure cloud services. We offer custom software, custom applications, business intelligence services, and artificial intelligence solutions for companies including AI for businesses, AI agents, and dashboards with power bi. Our team designs and integrates pipelines that combine 3D image processing, advanced ML models, and secure cloud deployment to accelerate return on investment.
Services proposal: Q2BSTUDIO can offer turnkey integration of the HCNN pipeline: µCT data ingestion, preprocessing, training, and deployment in aws and azure cloud environments, along with interactive power bi dashboards for result visualization and cybersecurity services to protect sensitive data. We also deliver custom software solutions that incorporate AI agents to automate control and analysis tasks.
Keywords and positioning: custom applications, custom software, artificial intelligence, cybersecurity, aws and azure cloud services, business intelligence services, artificial intelligence, AI for businesses, AI agents, power bi.
Conclusions: Automated prediction of granular material properties using multiscale convolutional analysis proves to be a viable and efficient path compared to traditional methods. The combination of 3D HCNN, µCT microtomography, and cloud deployment enables accelerating materials development and optimizing industrial processes. Q2BSTUDIO is prepared to transform this research into robust and secure commercial solutions, offering custom software and comprehensive artificial intelligence and cybersecurity services.
References: list of publications and µCT and mechanical testing protocols available upon technical request to Q2BSTUDIO.
