Additive manufacturing connects digital design, materials, and physical processes, but producing reliable parts still requires substantial human judgment. At MAIL, we develop AI methods that connect these stages: learning from sensor data to detect defects, using language models to retrieve manufacturing knowledge and plan corrective actions, and generating machine instructions from design inputs.

Our recent work includes LLM-3D Print, which uses visual feedback and language-model reasoning to monitor and adjust material-extrusion printing, and Image2Gcode, which uses a diffusion transformer to translate 2D images into printer-ready G-code. Together, these projects explore a more direct path from design intent to fabrication, with performance evaluated through physical printing experiments.

Image2Gcode: Image-to-G-code generation for additive manufacturing — Additive Manufacturing Letters, 2026.

We develop computer vision and multimodal sensing algorithms that continuously monitor the printing process, detect defects as they emerge, and provide real-time insights into build quality. By integrating optical, thermal, acoustic, and photodiode data, our models enable early anomaly detection and process understanding.

Representative Publications

LLM-3D print: Large Language Models to monitor and control 3D printing
Additive Manufacturing, 114, 105027 (2025).
Project and demonstrations · Code · CMU story (2026)

Deep learning based optical image super-resolution via generative diffusion models for layerwise in-situ LPBF monitoring
Additive Manufacturing, 107, 104790 (2025).

Additional Research Directions
Melt-pool monitoring, surface-defect classification, vision-based inspection, and photodiode-based process monitoring.

We are investigating how pretraining on manufacturing data can support reusable representations across printers, materials, and process conditions. This research direction focuses on reducing task-specific data requirements and adapting models to new additive manufacturing systems. Key questions include what knowledge transfers between machines, how much new data adaptation requires, and how reliably models perform under changing process conditions.

Printer Foundation Models

• Cross-machine representation learning

• Manufacturing foundation models

Pretraining for Manufacturing Intelligence

• Large-scale manufacturing representation learning

Transfer Learning Across AM Systems

• Domain adaptation for new materials and machines

We develop manufacturing-specific large language models that assist engineers in process planning, parameter optimization, failure diagnosis, and knowledge retrieval. These AI copilots combine manufacturing expertise with reasoning capabilities to accelerate design, troubleshooting, and decision-making.

Featured Work: AMGPT
AMGPT combines a pretrained language model with retrieval from additive manufacturing literature to support domain-specific questions.

AMGPT: a Large Language Model for Contextual Querying in Additive Manufacturing
Chandrasekhar et al. — arXiv preprint (2024).