The relentless scaling of silicon complementary metal-oxide-semiconductor (CMOS) technology, the engine of computational progress for over five decades, is approaching fundamental physical limits as transistor dimensions shrink below 5 nanometers. Concurrently, the computational demands of artificial intelligence (AI), which have been doubling approximately every 3.4 months, are vastly outpacing the semiconductor industry’s capacity to deliver performance gains through conventional silicon scaling alone. This widening gap between AI’s computational appetite and silicon’s physical ceiling represents a critical challenge for the continued advancement of computing. This dissertation investigates whether carbon nanotube (CNT) and two-dimensional (2D) material transistor architectures can overcome silicon’s physical boundaries to meet the performance, energy efficiency, and reliability requirements of next-generation AI microprocessors. Employing a quantitative, non-experimental research design, the study analyzed 5,742 compounds across four material families—transition metal dichalcogenides (TMDs), carbon-based materials, III-V semiconductors, and silicon—drawn from the Materials Project database, integrated with MLPerf benchmark data and OpenML evaluation results. The investigation proceeded through four interconnected phases: materials property analysis, AI hardware benchmarking, materials informatics via machine learning, and cross-phase comparative assessment. Key findings revealed statistically significant differences in electronic properties across material families (Kruskal-Wallis H = 1108.33, p < 2.1 × 10⁻²³⁷, η² = 0.192), with TMDs demonstrating the highest semiconductor fraction (56.2%) and favorable thermodynamic stability (formation energy −0.80 ± 0.34 eV/atom). Machine learning models achieved moderate predictive accuracy (formation energy R² = 0.73; band gap R² = 0.47; metal classifier accuracy 84.2%, AUC = 0.912), partially supporting the hypothesis of within-10% prediction accuracy. Literature-based analyses supported hypotheses regarding contact resistance below 100 Ω·μm for CNFETs and radiation tolerance exceeding 100 Mrad(Si). The findings contribute to semiconductor physics, materials informatics methodology, and national technology policy by providing an integrated, data-driven assessment of post-silicon material candidates within the strategic context of the CHIPS and Science Act.
Laszlo Pokorny (Tue,) studied this question.