Cross-Attention Fusion of Scientific Text and Project Metadata for Explainable Technology Readiness Level Prediction

Abstract

Technology Readiness Level (TRL) assessment is often constrained by document volume, evaluator subjectivity, and variation in institutional reporting practices. This paper presents a multimodal framework for TRL prediction that combines scientific project descriptions with structured project metadata through a cross-attention fusion layer. The dataset contains 18,247 technology projects collected from NASA TechPort, EU CORDIS, and US DOE repositories. NASA records provide explicit TRL labels, while CORDIS and DOE labels were derived using a milestone-based TRL rubric and annotator agreement analysis (κ=0.72). Textual descriptions were encoded with SciBERT, and structured variables were transformed through an XGBoost-based metadata encoder before being aligned in a cross-attention module. Evaluation was conducted using stratified 5-fold cross-validation with source-aware grouping to reduce leakage between related records. The proposed model reached 71.8% accuracy (MAE=0.82 TRL levels), 68.4% macro F1, and 89.2% adjacent accuracy, compared with 64.2% accuracy for the text-only SciBERT baseline and 54.1% accuracy for the structured-only XGBoost baseline. SHAP analysis identified project duration, funding amount, and technology domain as influential features, although these associations should be interpreted as predictive rather than causal. A trajectory extension produced 67.3% accuracy for 6-month TRL forecasting, with lower performance at longer horizons (62.8% at 12 months, 58.1% at 24 months). These findings support the use of multimodal TRL prediction as a decision-support mechanism, while acknowledging limitations related to label noise, cross-source bias, and external validation requirements.