Refine
Document Type
- Article (4)
Language
- English (4)
Is part of the Bibliography
- yes (4)
Keywords
- Architecture (1)
- Building and Construction (1)
- Civil and Structural Engineering (1)
- autoencoder (1)
- machine learning (1)
- neural network (1)
Institute
Material development processes are highly iterative and driven by the experience and intuition of the researcher. This can lead to time consuming procedures. Data-driven approaches such as Machine Learning can support decision processes with trained and validated models to predict certain output parameter. In a multifaceted process chain of material synthesis of electrochemical materials and their characterization, Machine Learning has a huge potential to shorten development processes. Based on this, the contribution presents a novel approach to utilize data derived from Small-Angle X-ray Scattering (SAXS) of SiO_2 matrix materials for battery anodes with Neural Networks. Here, we use SAXS as an intermediate, high-throughput method to characterize sol–gel based porous materials. A multi-step-method is presented where a Feed Forward Net is connected to a pretrained autoencoder to reliably map parameters of the material synthesis to the SAXS curve of the resulting material. In addition, a direct comparison shows that the prediction error of Neural Networks can be greatly reduced by training each output variable with a separate independent Neural Network.
In the context of environmental protection, the construction industry plays a key role with significant CO2 emissions from mineral-based construction materials. Recycling these materials is crucial, but the presence of hazardous substances, i.e., in older building materials, complicates this effort. To be able to legally introduce substances into a circular economy, reliable predictions within minimal possible time are necessary. This work introduces a machine learning approach for detecting trace quantities (≥0.06 wt%) of minerals, exemplified by siderite in calcium carbonate mixtures. The model, trained on 1680 X-ray powder diffraction datasets, provides dependable and fast predictions, eliminating the need for specialized expertise. While limitations exist in transferability to other mineral traces, the approach offers automation without expertise and a potential for real-world applications with minimal prediction time.