Delivering Predictions of Manufacturing Process Outcomes in Seconds, HP Integrates Artificial Intelligence Solutions to Control 3D Printing Deformation Challenges

Delivering Predictions of Manufacturing Process Outcomes in Seconds, HP Integrates Artificial Intelligence Solutions to Control 3D Printing Deformation Challenges

According to market insights from 3D Science Valley, leading companies have integrated advanced AI solutions into their workflows. For example, previously, Foxconn Group implemented 

AI for automated high-precision inspection of its product components and tools using NVIDIA software libraries and the NVIDIA EGX platform for accelerated computing.

According to Dr. Jun Zeng, Distinguished Technical Specialist for 3D Printing Software and Digital Twin Group Leader at HP, HP's team has been developing physical simulation engines 

based on first principles, calibrating these physical simulation engines with experimental sensing and metrology data in order to build on variations in the manufacturing process with the 

help of physical machine learning, which, once properly trained, can see orders of magnitude acceleration, and with models that can be run on laptops, this near real-time prediction 

provided by visible physics-ML machine learning opens the door to many new applications.

Driven by AI, 3D printing will pave the way out to be the main stream through self-evolving. 3D printing will become a mainstream manufacturing technology through self-evolving.

——Kitty Wang, Founder of 3D Science Valley

Artificial Intelligence for 3D Printing Defect Control - From Data to Prediction

© 3D Science Valley White Paper

Addressing the two key challenges that constrain the industrialization of 3D printing with respect to product quality: Predictability and Repeatability. Without solutions to these two key 

points, there is no point in simply pursuing the number of units sold.——Hewlett-Packard

Despite the unique appeal of additive manufacturing technology in achieving mass customized production and realizing complex designs, the application of this technology in 

manufacturing is still subject to a lot of resistance, adverse factors include: speed and quality of the final part or further investment is required to match the technology, enterprises for 

financial considerations, and so on. However, AI technology is stimulating the potential of additive manufacturing technology in key areas such as additive manufacturing design, process

development, quality control, and material development to drive the adoption of the technology in production.

The complexity of additive manufacturing design is interdependent with a multitude of factors, such as the quality of the material will affect the performance of the part, thus affecting 

the design decisions; production parameters will affect quality assurance, and quality assurance requirements will be reflected in those design decisions ...... and so on.

In the face of such great design complexity, the industry should think more about the problem is not how to utilize AI in additive manufacturing, but if there is no AI-driven design, 

production, quality assurance process, only by the power of human designers and engineers, we can still take advantage of additive manufacturing technology to improve product 

performance, accelerate innovation and other aspects of the advantages.

More versatile and efficient

The open ecosystem of Physics-Informed Machine Learning (physics-ML) fosters innovation in AI engineering applications. Physics-Informed Machine Learning embeds knowledge of the

laws of physics that control a given dataset into the learning process. This allows scientists to utilize prior knowledge to help train neural networks, making them more versatile an

efficient.

However, since physics-ML machine learning is an evolving field of research, domain experts need a better starting point to understand how it can be applied to their real-world use cases

. NVIDIA NVIDIA Modulus, an open-source framework for building, training, and fine-tuning physics-ML machine learning models using a simple Python interface, provides reference app

lications to fulfill the need. In this regard, the HP-Hewlett-Packard Digital Twin team found NVIDIA Modulus to be an ideal open innovation platform to contribute their work to support 

and collaborate with the broader manufacturing community.

Highlight- PIML

Physics-Informed Machine Learning (PIML) is an approach that combines a priori knowledge of physics with data-driven machine learning models. This approach effectively alleviates the problem of training data 

shortage, improves the generalization ability of the model, and ensures the physical soundness of the results.PIML has demonstrated its strong application potential in several fields, such as computational fluid 

dynamics, structural mechanics, and computational chemistry.

Physics-Informed Neural Networks (PINN), as an important branch of PIML, make it necessary to not only fit the data, but also satisfy the physical laws during the training process by encoding the physical laws in 

the form of Partial Differential Equations (PDEs) into the loss functions of neural networks.PINN models are usually composed of deep neural networks and are characterized by the inclusion of physical information

terms in the loss function, e.g., in fluid dynamics the Navier-Stokes equations may be used as physical information.

The steps to construct a PINN model mainly include:

Define the problem domain and the corresponding physics laws.

Designing a suitable neural network architecture.

Prepare the dataset, although PINN is less data dependent, data is still important for model training.

Define the loss function, which usually contains a data error term and a physical information error term.

Train the model using an optimization algorithm to minimize the overall loss.

Validate and test the model to ensure that it generalizes and conforms to the laws of physics.

Adjust hyperparameters and network structure to optimize model performance.

Interpret model predictions and apply them to real-world problems.

The main difference between PINNs and traditional machine learning models is that PINNs incorporate physical constraints during the training process, which allows them to give physically intuitive predictions 

despite low or noisy data. In addition, PINNs are more generalizable, making them particularly suitable for scientific computing and engineering problems that can be described by explicit physical laws.

NVIDIA Modulus is a physics-based machine learning framework that supports the construction, training, and fine-tuning of deep learning models of physical systems, also known as Physics ML models.Modulus 

supports new network architectures including Graph Neural Networks (GNNs) and Recurrent Neural Networks (RNNs) and facilitates the development of research into enterprise-grade solutions through open 

source collaboration.

The research area of PIML is rapidly evolving and covers a wide range of aspects from alternative model simulation, data-driven PDE solvers, parameterization of physical models, downscaling models, to knowledge

 discovery.Physical knowledge in PIML includes classical mechanics, symmetries and invariants, numerical methods for partial differential equations, and Koopman's theory, which can be augmented by data 

augmentation, neural network architectural design, and optimization of physical information, among other methods for integration into machine learning models.

In summary, PIML and PINN provide a new perspective and tools for solving traditional scientific computing and engineering problems, and they improve the prediction accuracy and generalization ability of models 

by combining physical knowledge and data-driven learning. With the deepening of research and the development of open-source tools, the application of PIML will be more promising.

Breaking through speed constraints

Traditional high-fidelity physics simulation workflows are computationally intensive, with a single design iteration typically taking hours to days to complete. Using low-fidelity, 

reduced-order models can severely limit design exploration. Physics-ML machine learning alternative models provide high-fidelity simulation and complement numerical solvers to 

increase the speed of design iterations by several orders of magnitude.

For example, Physical-ML Machine Learning Alternative Models can now be used to provide immediate feedback on the manufacturability of product designs and automated design 

filtering through large design spaces to optimize functionality and yield. Filtered designs can be simulated in greater detail using numerical solvers. These AI models also enable product 

design teams to use their previous simulation databases as a source of real data.

Today, different engineering departments are responsible for product design, which optimizes for functional attributes, and product manufacturing, which optimizes for yield. The final 

product design optimized for both requires multiple iterations between the two engineering departments, which can take weeks to months. This is a significant bottleneck for new product

launches.

Breaking through the speed constraint, 3D Science Valley has learned that HP is developing a digital twin for its Metal Jet metal 3D printing technology that enables process engineers to 

predict and optimize design parameters and process control parameters to improve part quality and manufacturing yields.

The diagrams show the different stages of Metal Jet printing: metal powder layup, binder jetting, billeting, shelling, sintering, cooling, and finishing. The two photos below show the inputs 

and outputs of the HP Metal Jet.

Figure 1. Simulating the complex metal sintering process in HP Metal Jet printing is critical to optimizing yield

——© HP-Hewlett-Packard

For example, as part of the HP Digital Twin effort, the HP team developed the Virtual Foundry Graphnet model through applied physics-ML machine learning to significantly accelerate 

calculations for predicting phase transitions in metal powder materials. This trained alternative model has achieved orders of magnitude acceleration, enabling near real-time, high-fidelity 

simulation of the metal sintering process.

Virtual Foundry Graphnet has also demonstrated that such AI alternative models can be applied to designs with varying geometric complexity and different configurations of process 

parameters. For example, a dragon model with mesh overlays shows process-induced distortions due to volume shrinkage, gravity sag, collapse, bending, and friction effects.

▲Figure 2. Stanford Dragon Test Model

——© HP-Hewlett-Packard

The Stanford Dragon test model in Figure 2 emphasizes the need for simulations that take into account computational materials engineering and the physics of the manufacturing process

to accurately predict the geometric deformation of the final part caused by the manufacturing process.

Physics-ML Innovation at HP

The HP Digital Twin Digital Twin team believes that open source communities play an important role in accelerating the development of Physics-ML machine learning and expanding its 

applications, and NVIDIA Modulus provides an excellent platform to help and support such open source communities. HP 3D Printing has joined the physics-ML open source community by 

open sourcing Virtual Foundry Graphnet through the NVIDIA Modulus platform. However, Physics-ML machine learning and its application to real-world materials engineering problems is 

still in the early stages of industrial adoption, and more research is needed to extend these methods to a variety of use cases.

With physical-ML models such as Virtual Foundry Graphnet, engineers can co-design for functional yield and significantly accelerate time to market. Currently, 3D Science Valley understands 

that Digital Sintering, HP's process physics simulation software, has been deployed in real-world HP Metal Jet application scenarios to improve manufacturing results. Digital Sintering 

generates improved designs that can compensate for part distortion caused by the manufacturing process.

Two images show a design generated by Digital Sintering after a physical manufacturing process that produces a fabricated metal object with geometric accuracy.

Figure 3. Improved design to compensate for part distortion caused by the manufacturing process

——© HP-Hewlett-Packard

Running a trained metal sintering inference engine takes only a few seconds to obtain the final sintered deformation value. Figure 4 shows a 63 mm test part with a maximum nodal error 

within 2%. The complete sintering cycle takes about 4 hours. The average difference between the predictions made by the Physical-ML machine learning and the predictions generated by 

the physical simulation in this case is 0.3 mm.

Overall, Physics-ML machine learning is at the forefront of near-real-time simulation workflows, and HP 3D Printing's Physics-ML machine learning innovations (e.g., Virtual Foundry 

Graphnet) demonstrate the power of AI to dramatically accelerate simulation workflows and provide predictions of manufacturing process outcomes in seconds.

Knowing is as deep as it is far-reaching. Based on a global network of superb manufacturing experts and think tanks, 3D Science Valley provides the industry with an in-depth look at 

additive and smart manufacturing from a global perspective. For more analysis on the additive manufacturing space, follow the white paper series published by 3D Science Valley.