Machine learning accelerated discovery of chalcogenides for memristors
Project Director: Dr. Alin VELEA
Project: 60PCE din 08/01/2025 (PN-IV-P1-PCE-2023-1785)
Contractor: National Institute of Materials Physics (NIMP)
Project leader: Dr. Alin Velea
Project type: Program 5.1 - Ideas, Exploratory research projects (PCE)
Start date: 08.01.2025
End date: 31.12.2027
Funding Agency: UEFISCDI
Abstract:
GlassyMEM aims to accelerate the discovery of chalcogenide glasses for memristor applications through the use of machine learning methods. Chalcogenide phase change materials possess excellent electrical properties, including high resistivity contrast between glass and crystalline states and dynamic resistance tuning, making them ideal for memristors in AI, big data, and IoT applications. Traditional approaches for identifying new chalcogenide glasses rely on empirical rules, lacking a universal predictor of glass formation ability (GFA).
In this project, we will construct a comprehensive database integrating literature and experimental data obtained within this project. Chalcogenide libraries will be synthesized via magnetron co-sputtering, and their structural and compositional properties will be extensively characterized. Machine learning models will be developed and trained to predict GFA for chalcogenides, enabling faster and more efficient material screening.
Furthermore, the electrical properties of the most promising candidates will be investigated for memristor applications, aiming to achieve more powerful, energy-efficient, scalable, and cost-effective memristor devices. By adopting a data-driven approach, GlassyMEM will contribute to a deeper understanding of GFA, reducing costs, risks, and time associated with conventional methods. The project will propose viable candidates for advanced memristor technologies.
Project Leader: Dr. Alin Velea - Scientific Researcher I
Research Team:
Dr. Florinel Sava - Postdoctoral Student
Dr. Iosif Daniel Simandan - Postdoctoral Student
Dr. Claudia Mihai - Postdoctoral Student
Dr. Amelia Elena Bocirnea - Postdoctoral Student
Dr. Mohamed Yassine Zaki - Postdoctoral Student
Dr. Adelina Udrescu - Postdoctoral Student
Angel-Theodor Buruiana - PhD Student
Miruna-Ioana Belciu - PhD Student
Ciprian-Augustin Parloaga - Master Student
Main Results Obtained
Year 1 (2025) Achievements
1. Database Development
We are building a comprehensive chalcogenide database by:
- Extracting and processing data from the SciGlass database
- Manual digitization of glass formation data from scientific literature
- Coverage of binary, ternary, and multicomponent chalcogenide systems
This database serves as the foundation for training and validating our machine learning models.
2. High-throughput experimental work:
- Synthesis of combinatorial thin film libraries by magnetron co-sputtering
- Structural characterization using X-ray diffraction (XRD) and X-ray reflectivity (XRR)
- Compositional analysis using Rutherford Backscattering Spectrometry (RBS)
- Optical characterization by spectroscopic ellipsometry
These experimental activities provide data for model validation and exploration of new chalcogenide compositions.
3. Machine Learning for Optical Property Prediction
We developed and published an ensemble machine learning approach for predicting the refractive index of chalcogenide glasses:
- Trained and compared 8 base models (RandomForest, AdaBoost, CatBoost, XGBoost, HistGradientBoosting, LightGBM, ExtraTrees, KNeighbors)
- Created ensemble meta-models (Voting and Stacking classifiers)
- Computed physicochemical descriptors based on elemental properties
- Applied SHapley Additive exPlanations (SHAP) for model interpretability
- Identified key features governing optical properties in chalcogenide systems
This work demonstrated the effectiveness of ensemble machine learning methods for predicting chalcogenide glass properties.
Article: M.-I. Belciu, A. Velea, "Ensemble Machine Learning for the Prediction and Understanding of the Refractive Index in Chalcogenide Glasses", Molecules 2025, 30(8), 1745, https://doi.org/10.3390/molecules30081745
Oral presentation: "Ensemble Machine Learning for the Prediction and Understanding of the Refractive Index in Chalcogenide Glasses", EMRS 2025 Spring Meeting (May 26-30, 2025, Strasbourg, France)
Cognitive and Socio-Economic Impact
Scientific/Cognitive Impact
Advancement of Materials Discovery Methodology:
- Demonstrated that machine learning can effectively accelerate chalcogenide materials discovery
- Building comprehensive chalcogenide databases for property prediction
- Provided interpretable insights through SHAP analysis, connecting ML predictions to physical mechanisms
Contributions to Fundamental Understanding:
- Identified key physicochemical descriptors governing optical and structural properties in chalcogenides
- Developing data-driven approaches for understanding glass formation
Training of Next-Generation Researchers:
- Training of Master students in interdisciplinary ML + materials science
- Developing expertise in high-throughput experimentation and data-driven materials discovery
Socio-Economic Impact
Technology Relevance:
Memristors are key components for neuromorphic computing, which promises:
- 100-1000× lower energy consumption than conventional computing
- Real-time AI processing at the edge (no cloud dependency)
- Brain-inspired computing architectures for complex pattern recognition
Potential Applications:
- Energy-efficient artificial intelligence hardware
- Internet of Things (IoT) sensor processing
- Autonomous vehicle perception systems
- Medical diagnostics and wearable devices
Sustainability:
- Reduced computational energy consumption addresses climate change concerns
- Chalcogenide materials use earth-abundant elements
- High-throughput screening reduces experimental waste and accelerates development
Dr. Alin Velea - alin.velea@infim.ro
PROJECTS/ NATIONAL PROJECTS
Copyright © 2026 National Institute of Materials Physics. All Rights Reserved