Feature published in the May–June 2020 issue of 3D ADEPT MAG, then online in September 2020. Roles, products and forecasts reflect that period.
How can artificial intelligence (AI) be used in additive manufacturing (AM)? This 2020 feature examines opportunities and limitations described by AMFG, Additive Flow and Ai Build, from design to process control.
AI encompasses different approaches. Machine learning is one subset: it uses data to develop models. The interviewees explain below how they use these terms.
“AI can be seen as an umbrella term, and its scope is incredibly vast. At AMFG, we see AI as a technology that will enable AM machines and systems to be ‘smart’ enough to carry out core operational tasks both effectively and efficiently.
But achieving this requires machine learning. Machine learning is a form of AI and, as the name suggests, it is the idea that machines can learn from data to solve problems and execute tasks – without manual intervention” states Keyvan Karimi, the CEO & Founder of AMFG, that we invited in this segment alongside Additive Flow’s CEO & Co-founder Alexander Pluke as well as Ai Build’s CEO and co-founder Daghan Cam.
If Pluke agrees with the fact that machine learning is a subset of AI inspired by scientific understanding of how humans learn, he outlines that he prefers “a broader definition of AI (and in this sense ‘narrow’ as opposed to a ‘general’ AI as depicted in sci-fi) – where a computational approach can deliver results or recommendations normally completed by human intelligence. In this sense, smart algorithms written by humans that can act ‘intelligently’ – for example evolutionary algorithms and generative approaches, as well as complex classical algorithms that pass the likely subjective threshold of ‘intelligence’, and substitute human activity, could be labelled ‘AI’”.
In 2020, the interviewees described an integration still developing across the AM workflow, with applications at different levels of maturity.
Their accounts present approaches considered at the time; they are neither a current market assessment nor independent validation of each solution.
Adapting to process changes
The first task is to define the industrial problem and the available data. This feature does not provide a representative measure of AI adoption in factories.
For Ai Build’s Daghan Cam, the problem with this uncertainty in the manufacturing landscape is that not only AI is misinterpreted but companies do not answer the right questions:
“AI is sometimes misinterpreted as a black box software that learns and solves any given problem. That is the definition of Artificial General Intelligence which doesn’t exist today, and most probably won’t exist in the next 10 years. What we have very commonly available instead is narrow forms of artificial intelligence which are extremely powerful in solving specific sets of problems.
In order to leverage these scientific breakthroughs, application developers in manufacturing like in any other industry must be crystal clear about two questions:
- What exactly are we trying to solve with AI?
- How are we going to source the specific type of data that is required to solve that problem?”
Well, it makes sense. How can we address a challenge if that challenge itself is not well defined? In other words, AI will remain a buzzword if there is no upstream strategy.
The applications discussed include quality control, production management and design support. NIST also studies data-driven support for AM design and process planning, with suitable methods and measurements.
Is AI a Must in AM?
AM selection depends on total cost, throughput, geometry and required properties. AI may assist selected operations; it does not replace comparison with other manufacturing methods.
The interviewees explore several potential uses of AI across this workflow.
First, AI allows for a new way of thinking about software.
“Any software driven technology that doesn’t use AI is doomed to fail and will be replaced sooner or later. There is a big responsibility on the machine manufacturers and software developers to work together on redefining the boundaries of additive manufacturing with new releases of their products in order to stay relevant . If a machine is able to bring practical or commercial value to its customers by using AI, it is highly unlikely that other competing products will be able to survive in the market without taking advantage of a similar feature within the next 5 to 10 years depending on the novelty of the process.
For example, Ai Build introduced computer vision into 3D printing in 2016 for autonomous path planning. At that time 3D printers didn’t have cameras embedded so we built a solution from scratch to develop our machine learning algorithms. Nowadays most industrial 3D printers are shipping with built in cameras because of the increasing demand from the market. This is great news because now we are able to work with innovative partners to develop computer vision solutions for different platforms and use cases. If we project 5 years forward, we don’t think any major industrial 3D printers will exist without basic computer vision capabilities like automatic fault detection, because the technology is too useful to be left out”, Ai Build’s CEO comments.
Historical context: these five- and ten-year forecasts are Daghan Cam’s opinions expressed in 2020, not established outcomes.
Opportunities across the workflow, from design, to production and Quality Assurance (QA).
According to Additive Flow’s Alexander Pluke, opportunities for AI extend from design to production and quality assurance.
“The complexity of the ‘design domain’ (the possible combinations of design outcomes within a set of variables) is both extremely large, and interdependent. Material quality will affect part performance, which will influence design decisions; and production parameters will affect quality assurance – and quality assurance requirements will (or should!) be reflected in those design decisions… and so on.
Bring in concepts like geometric freedom, and an Additive Flow speciality ‘multi-property’ – where material properties can be tailored within different regions of a geometry – and this complexity gets greater still.
In light of this – rather than asking the question ‘How can AI be leveraged in AM?’, the question should be ‘How can AM be leveraged without AI?’. This is more of a rhetorical question. Without deploying AI-driven tools (in design, in production, in QA, across the workflow), industrial use of AM will not be fully leveraging the opportunity (cost savings, performance, innovation acceleration) we have all heard about.”
Defect Detection
File analysis can help identify some geometric defects and suggest corrections before manufacturing. A corrected digital model alone does not guarantee a printable, conforming part: material, process and application requirements remain decisive.
“This enables humans to spend much less time on repetitive, manual tasks and more time on value-adding processes.
Machine learning is one step in this direction. For example, our software is able to provide intelligent material and technology recommendations based on previous data inputs. Our software can also predict part quality and allow users to control the 3D printing process so as to reduce the risk of printing errors. This saves both time and costs.
Additionally, as AM moves to the production of end-use parts, a vast amount of data will be needed to optimise processes. AI will play a role in leveraging this data to achieve faster production and optimal results, so that companies can manage the 3D printing process more effectively”, says Keyvan Karimi from AMFG.
Creating new materials with new properties
Process monitoring can help detect anomalies and reduce some scrap. Machine learning is also studied as a way to guide the search for materials and formulations meeting multiple objectives.
This involves proposing candidates using data and models, then evaluating them experimentally. Entering desired properties does not guarantee that they will be achieved, or establish cost or printability.
“Data, data, & data!”
These examples show potential synergies between AI and AM without making AI essential to every application. Their value depends on the problem, data quality and validation of the results.
However, for the three companies that take part in this segment, the most important challenge that slows down the integration of AI in AM is data.
“The best optimisations need the best data sets, and open access to data within this sector is not (yet) prevalent”, explains Pluke.
“One of the biggest challenges is the high level of complexity required to develop AI specifically for 3D printing. Large volumes of data are needed and extensive testing and experimentation done to ensure that the machine is ‘learning’ correctly.
Additionally, AM as an industry is still very much in the early days of integrating 3D printing and AI. Developments for these two technologies are largely happening separately”, adds Karimi.
How do the participating companies approach AI and AM?
The three companies interviewed for this feature were based in the United Kingdom: AMFG, Additive Flow and Ai Build.
AMFG presented software for managing the AM workflow. Its CEO and founder, Keyvan Karimi, describes the company’s development directions here.
The company described three directions for its manufacturing execution system (MES):
- File analysis and production-time estimation to prepare work and identify potential problems. This does not eliminate every risk of printing failure.
- Anticipating part quality using preparation and process data. This differs from predictive maintenance, which concerns equipment condition and servicing needs.
- Finally, “the third way in which we aim to use AI is through the automation of key process steps across the entire 3D printing workflow”, declares the company.
Additive Flow was developing an approach to optimising the digital AM workflow. Its CEO and co-founder Alexander Pluke presents the principles.
Its Formflow software aimed to use design and manufacturing data to allocate materials, properties and structures within a geometry.
“Some aspects of the design spaces we are presented with are best suited to neural network approaches, while others are best met with more classical algorithmic techniques. For example, using generative algorithms to optimise exponential variables allows our users to discover new high performing results by revealing solutions that would have otherwise been hidden. Neural networks have provided us some powerful results computationally, and have been useful when considering optimisations that rely on data-structures less digestible by classical algorithms,” explains Additive Flow.
According to Additive Flow, the aim is to use artificial intelligence to allocate materials, processing parameters and local structures alongside geometry, rather than adopting AI for its own sake.
Ai Build was developing AM software and hardware, including process control and inspection. Its CEO and co-founder Daghan Cam explains the approach in this feature.
The company first came to our radar two years ago when its technology was displayed in the Sculpture gallery of the Victoria and Albert Museum. Over time, we realized it also stands out from the crowd thanks to its AiSync software.
“One of the foundational decisions we made in AiSync software to enable an AI driven additive manufacturing process is to deliberately avoid using gcode files to describe a toolpath. This was very counterintuitive because all 3D printers and CNC machines on the market that we are aware of use gcode or a brand specific gcode-like language to drive the motions of machines. Gcode would work fine in a perfectly deterministic world where the whole process could be predicted with high accuracy, but unfortunately the additive manufacturing process in real life is far from perfect. We are not able to predict the physical form of 3D printed parts precisely because even the tiniest change in the build volume during production can lead to catastrophic deviations in the additive process. Even if we could fully control the environment and if there was a powerful simulation engine to precisely predict material behaviour 99.9% of the time in one layer, the success rate of the overall prediction on a typical part with thousands of layers that support each other would still be less than 30% for statistical reasons.
This was the key insight for us to abandon gcode and develop a completely new machine control process from scratch that is more resilient. In AiSync software we use a combination of offline and online optimization methods to control the actions of 3D printers. Instead of slicing a design and uploading static gcode files onto machines, AiSync analyzes a design with powerful computers on the cloud and sends optimized directives to the machines in an abstract format. These directives are then interpreted and turned into machine level instructions bit by bit on the fly by another computer that is located on the edge which has real-time access to sensor data from the machine. This double optimisation infrastructure enables us to take full advantage of AI algorithms by combining the strengths of cloud supercomputing for high level tasks like path planning and edge computing for time critical tasks like fault detection and quality assurance”, describes the company.
Editorial clarification: the probability example above is illustrative. Calculating success across layers requires specifying their number and dependencies between events; it is not a general measurement of 3D-printing reliability.
To sum up…
This 2020 feature highlights the importance of data, access to it and collaboration. Integrating AI requires specific objectives, testing and process-appropriate validation; the accounts do not demonstrate a universal solution.
This dossier has originally been published in the May–June 2020 Issue of 3D ADEPT Mag.
THE 3D ADEPT WEEKLY NEWSLETTER
The industry moves fast. Stay ahead.
Every week, the latest additive manufacturing news, in-depth features and real-world applications, straight to your inbox. Plus, be the first to know when a new issue of 3D ADEPT Mag is out.
Free · One email a week · Unsubscribe anytime
Clicking loads our email provider’s sign-up form, including its analytics and Google’s anti-bot protection. You’re only subscribed if you complete the sign-up. Read our privacy policy.