Advanced Informatics and Artificial Intelligence for Additive Manufacturing

Summary

Advancements in additive manufacturing (AM) are increasingly driven by advanced sensors and measurements informing increasingly complex modeling and simulation paradigms. Such advanced informatics are providing new opportunities to harness trusted data and information to acquire knowledge and develop actionable assessments in complex AM systems and environments, facilitating part design, production and qualification. 

The Advanced Informatics and Artificial Intelligence for Additive Manufacturing (AI2AM) project aims to develop the methods, models, standards, and best practices to reduce unknowns in AM part fabrication and qualification towards “born qualified” and “first part correct” outcomes.  The AI2AM project leverages technologies such as machine learning, digital twins, digital threads, and artificial intelligence models to promote process assurance and quality assurance during AM part fabrication.

The complex systems addressed in this project extend beyond the design-to-part transformation, allowing for considerations in production, operations management, supply chain agility, and end use scenarios. This holistic approach will support industry, including SMEs and LSIs, in efforts to enhance AM capabilities and fully integrate AM technologies in meaningful, beneficial use cases.

Description

Credit: NIST

Objective
To develop and deploy the metrics, models, reference data, and best practices for implementing and adopting advanced informatics (including product definition, digital twins, ontology, and machine learning) and artificial intelligence to enhance additive manufacturing design, process planning, and fabrication with the intent of reducing lead times, supporting first-part-correct goals, and exploiting the inherent advantages of additive manufacturing processes.

Technical Idea
The AI2AM project operates with three primary objectives:

1. Enhance the capabilities and experience of AM practitioners by facilitating, testing and validating the application of emerging AI and ML technologies throughout the design-to-product-transformation,
2. Improve the predictive analytic capabilities for first-part correct design and manufacture and reduce the uncertainties and unknowns associated with the specification, qualification and acceptance of an AM part,
3. Inform AM-enabled supply chains and facilitate viable, robust supply chain alternatives by eliciting dependencies at process, part, facility, enterprise and supply chain levels of abstraction.

To achieve these objectives, the AI2AM project will leverage domain expertise and information sciences.

To increase success rates and facilitate the adoption of AI and ML techniques, the models that drive them must be customized for specific problems and problem types. This customization requires benchmarks and metrics to establish baselines and evaluate against. Establishing baselines and best practices for the adoption of tools founded in ML and AI technologies will provide AM practitioners with much needed insight into their applicability, utility, and practicality.  Further domain-driven adjustments and enhancements of AI foundational models will be facilitated by establishing resources and best practices for hyperparameters, fine tuning, artificial narrow intelligence, prompt engineering, and retrieval-augmented generation (RAG).

Digital twins provide the ability to digitally “manifest” a physical part at various stages of the transformation and with high rates of observation.  Accompanied by deliberate, precise analytics, digital twins have the potential to provide previously unachievable quality assurance in a digital manufacturing process.  Advances in predictive analytics will include addressing uncertainties through the development of new methods to quantify the fidelity of individual models and facilitate their continued reuse, including model aggregation and composition. Advances in analytics, coupled with sector-specific considerations for accompanying digital twins, will open new pathways to quality assurance of AM parts. The project will develop a framework for the adoption of comprehensive information models, digital twins and associated analytics to support the continuous specification and validation of an AM part throughout the design-to-product transformation.

Scaling digital twins beyond the part, to the machine, facility, enterprise, and supply chain can provide newfound transparency into AM industrialization scenarios. Methodically developed and incorporated, these multi-scale digital twins (aggregates) can provide uninhibited access to understanding and mapping the integration of AM technologies in the creation of robust, AM-supported supply chains. This transparency can be used to both increase confidence in the technology and expedite adoption and acceptance.  The development of digital twin aggregates, case studies and best practices in this area will alleviate the initial burden on SMM adoption of AM technologies.

Research Plan
The AI2AM project takes a synergistic approach to addressing the three problems and objectives previously outlined.  The methods, models, analytics, and best practices developed by the project will complement each other (see Figure) so as to build a shared foundation on which goals can be achieved and problems overcome.

The foundation of the project will be developed on improved information structure, both in specifications and observations.   Efforts in data packages, product definition, and Verification, Validation, and Uncertainty Quantification (VVUQ) will strengthen the correlation between how requirements are specified and satisfied.  To support data packages and product definition, the project will develop methods and models for correlating physical parts, at different stages of maturity, to part specifications, at multiple scales (micro, meso, macro) to facilitate product specification and in situ process and part assurance.  To develop VVUQ methods, the AI2AM project look to develop a strategy and framework for implementing VVUQ to support the quantification of measurement uncertainty in AM. The AI2AM project will explore how current best practices in various domains can be adopted to support VVUQ specifically for AM, including the unique challenges stemming from the many variations of process physics, sensor configurations, and parameter combinations.  The project will engage relevant efforts in model validation and standards development while initiating new efforts as needed.  

Towards the development of improved “first part correct” goals, and process and quality assurance for intended applications, the AI2AM project will develop the methods, models, and representations necessary to realize robust digital twins of AM processes and parts.  A “fit-for-purpose” approach to the development of digital twins, including the development of domain and application-specific digital twin models that integrate specific acceptance requirements into the design and process planning.  The AI2AM project aims to develop the methods and models necessary to create aggregate digital twins, customizable to an AM part and process, and the methods to establish qualification thresholds with these digital twins.  These digital twins will be supported with development of improved predictive analytic capabilities, including 1) new methods and standards to quantify the fidelity of AM models and simulations through VVUQ and 2) new methods and best practices for leveraging AI and ML tools and building trust into their application.

Predictive analytics are increasingly realized through the adoption of ML and AI techniques.  To support the adoption of ML and AI techniques in AM, AI2AM will develop methods, models, and best practices for their application. The project will explore and characterize the behaviors of various approaches such as neural networks and large language models (including those associated with foundational models) towards the establishment of best practices for developing training data and managing unintended bias. The use of AM-derived ontologies will be explored as a means for introducing and maintaining context when training various models for AM applications, including the development of data-driven design rules.  The project will curate and host a set of resources on which other AI models can be trained and customized to solve AM-specific problem sets.

Towards the development of robust AM-enabled supply chains, the AI2AM project will: 1) Define and develop the methods and models necessary for realizing agile, multi-scale digital twins for supply chain integration and 2) Define and develop the methods and best practices for securing supply chain integrity.  Integration of part and system-level digital twins into the larger enterprise requires characterization of AM processes within a production environment, including build times, preparation times, and post processing times.  To inform enterprise-level decision making, these part and system digital twins must be normalized to be interoperable with digital twins of traditional production environments, including defining multi-scale interactions. To support transparency through digital twins, new methods and communication protocols are needed to convey the state of a part at different stages of fabrication.  To support production agility, new system-level digital twins, including machine emulators, are necessary to support machine selection and reconfiguration.

Major Accomplishments 

• Methods for Data packages
• Standard for Product Definition
• Models for advanced simulation and path planning
• Reference models for AM  digital thread and digital twin

Recent Accomplishments

• Completion of ASME Y14.46 (FY22) and subsequent by MIL STD 31000C (FY25)
• CD Ballot of AP328 E4 Extension Numerical Control for PBF (To DIS in FY26)
• ASTM Standardization/ DIS Ballot of “Data Packages for AM Parts” (To FDIS Ballot in FY26)