실험–계산 통합을 통한 광발전 재료의 광전자 특성 예측을 위한 인공지능 프레임워크 = Artificial Intelligence Framework for Optoelectronic Property Prediction in Photovoltaic Materials through Experiment–Computation Integration

다국어 초록 (Multilingual Abstract)

Photovoltaic technology stands at a critical juncture: while solar energy represents our most abundant renewable resource, discovering the next generation of high-performance photovoltaic materials remains prohibitively slow and expensive. The fundamental challenge is staggering in scale millions of potential molecular structures and compositional combinations exist within the organic semiconductor and perovskite material spaces, yet traditional approaches can evaluate only a tiny fraction. Experimental synthesis and characterization require months per candidate material, while quantum mechanical calculations, though more feasible, still demand computational resources that render systematic screening of vast chemical spaces impossible. This thesis transforms this paradigm by establishing a comprehensive artificial intelligence framework that bridges quantum mechanical accuracy with high-throughput screening capability, enabling researchers to predict critical optoelectronic properties across diverse photovoltaic material classes in milliseconds rather than days.
The framework's development follows a strategic progression through the photovoltaic materials landscape, each investigation building upon previous insights while addressing increasingly complex challenges. Beginning with organic materials the foundation of molecular photovoltaics the research develops graph attention networks that learn to predict electronic band gaps by identifying which molecular features most strongly influence energy levels. Training on 11,369 molecules with quantum mechanical property calculations, these models achieve predictions approaching DFT accuracy (mean absolute error 0.26 eV) while reducing computational cost ten thousand fold, demonstrating that artificial intelligence can extract the underlying physics governing structure property relationships. This success with fundamental electronic properties creates the foundation for the next challenge: organic semiconductors, where practical device operation depends not on a single property but on the interplay of multiple electronic characteristics. Here, the framework evolves to employ residual gated graph neural networks capable of simultaneously predicting frontier molecular orbital energies, excitation energies, and oscillator strengths across 48,182 compounds. The ability to capture multiple interdependent properties through unified architectures proves crucial, achieving accuracies exceeding 84 correlation with quantum calculations while revealing how molecular structure simultaneously controls light absorption, charge injection, and excited-state dynamics. The investigation then extends beyond molecular materials to perovskite solar cells, where compositional engineering of crystalline structures offers unprecedented tunability but introduces fundamentally different modeling challenges. Adapting the framework to handle categorical compositional variables rather than molecular graphs, specialized machine learning models successfully predict both electronic properties and device-level performance metrics directly from composition, achieving 84% accuracy for band gaps and 63% for power conversion efficiencies—remarkable given the complexity of translating composition through electronic structure to device performance.
What unifies these diverse investigations into a coherent framework is not merely the application of machine learning to different material classes, but rather a rigorous methodology for ensuring that computational predictions translate reliably to experimental reality. Each model is trained on large-scale computational databases that provide systematic, extensive coverage of chemical space, then validated against experimental measurements from synthesized materials reported in literature. This experiment-computation integration proves essential: models validated only against additional computational data may achieve impressive metrics while failing to predict actual material behavior, but systematic validation against 13-187 experimental samples per material class confirms that the framework's predictions reliably guide experimental materials selection. The framework thus achieves what neither purely computational nor purely experimental approaches can accomplish alone the systematic exploration of vast chemical spaces with accuracy sufficient for practical materials design.
The practical realization of this framework through accessible computational tools represents a crucial contribution beyond methodological advances. Web-based prediction servers and open-source implementations democratize these capabilities for researchers worldwide, eliminating barriers of computational expertise or infrastructure access that might otherwise limit adoption. A materials scientist can now input molecular structures or perovskite compositions and receive property predictions in seconds predictions that would require weeks of laboratory work or days of supercomputer time through traditional methods. This transformation in the economics and accessibility of property prediction enables fundamentally new approaches to materials discovery: systematic screening of millions of candidates becomes feasible, inverse design workflows can identify structures matching target property profiles, and researchers can rapidly evaluate hypotheses about structure-property relationships that would be impractical to test experimentally. By establishing artificial intelligence as a practical, validated tool for photovoltaic materials discovery rather than merely a theoretical possibility, this work accelerates the development of next-generation solar technologies that are essential for addressing global energy and climate challenges.

목차 (Table of Contents)

• 1 Introduction 1
• 2 Artificial intelligence and datasets for computational materials science 19
• 3 Residual gated RGNNS for organic semiconductor optoelectronic properties 47
• 4 Graph attention networks for organic materials bandgap prediction 79
• 5 Explainable machine learning for perovskite bandgap and efficiency 95
• 6 Conclusion and Future Direction 120