Beyond ChatGPT: How machine learning is delivering productivity gains in additive manufacturing

ChatGPT has thrown the benefits of Artificial intelligence (AI) into sharp relief. In just two months, it got to 100 million users. Online and in the media, the public is sharing examples of using the chatbot to automate work, check code and create content outlines. It also triggered an avalanche of new products offering AI capabilities. 

Does all of this amount to a significant technological leap? Has AI become a genuinely useful tool or is this just another peak of a hype cycle, the likes of which we’ve seen in the past? 

ChatGPT is just one of many innovations coming out of the developing field of AI that really reached a tipping point in the last decade.  

At Oqton, we’ve been applying machine learning models to automate manufacturing. We’ve been an AI company since day one, and we have a dedicated AI team who looks into using AI smartly, in areas where it can make a real difference.  

When I started the company, five years ago, I believed that we were on the brink of AI becoming a tool for achieving productivity gains – and not just a gimmick. That’s exactly what we’re witnessing today. 

Beyond ChatGPT, the level of AI capabilities that we can now reliably build will transform every part of our economy. Manufacturers using additive technologies are already reaping huge rewards from Oqton’s AI.  

Oqton has been working with organizations using a range of technologies, from robotics, to 3D printing and CNC milling, in the dental, medical and industrial sectors. The general principles behind our AI and the newer kids on the block, like ChatGPT and Bard, are similar. Let's take a look at what they are. 

What are AI, machine learning and neural networks? 

In simplest terms, machine learning is a subset of AI that uses algorithms trained on data to produce models that can perform complex tasks. Strip away all the hype, and it boils down to statistics on steroids. Think of a non-linear function with 200 billion parameters instead of just two. 

The workhorse of the most popular machine learning applications today, like ChatGPT, OpenAI’s image generator DALL-E, and self-driving cars, are neural networks. This type of machine learning model consists of neurons – a statistical algorithm invented around the 1960s. The name “neuron” comes from the limited knowledge people had back then about how the brain worked.  

Neural networks are especially good with ill-defined patterns in data such as speech recognition, image recognition, and understanding of the context in written language. Wherever things are vague, noisy or not precisely defined, that's where neural networks  shine. 

In the field of machine learning you hear the term deep neural network or deep learning. Around 2010 UK researcher Geoffrey Hinton came up with the idea to have many layers of neural networks, from a dozen or down to 50 layers of neurons. 

Scientists found ways of efficiently training such a neural net, and learnt that these had very powerful representation capabilities. In essence, they found it was much more efficient to make neural networks deep than to make them wide. 

Another important factor for the advancement of machine learning is computing power. Models need to process enormous volumes of information. 

Computer scientists were able to achieve this only in the last couple of decades. They leveraged GPUs that were originally built for graphics rendering and discovered they could also be very capable architectures to efficiently optimize very large mathematical functions. 

But the trick is to take this function, which could do anything, and build a data set for a high value domain such as manufacturing. 

Forget the algorithms. It’s the data sets. 

The basic principle of machine learning is that you train an algorithm on data sets. They include examples, numbering in thousands and thousands, or millions, to which we add labels. For example, annotating an audio recording of spoken language is labelling. 

You give the data set to an algorithm and it uses the knobs for your function to approximate the data in the best possible way. 

What’s been shown is that, when data sets are large enough, these techniques can generalize. The algorithm is trained on one data set and it learns the structure of that data. You can then apply it to unseen data points and get good predictions. 

In AI, the data sets make all the difference. The algorithms themselves are all open-source, and the main research labs in the top companies like Facebook, Google, and Microsoft, publish their work with code. Consequently, everyone has the same algorithms. What really matters is how much data you have and what the quality of the data is. 

And of course, whether you can process it. Training ChatGPT costs about $10M in electricity bills. Thousands and thousands of GPUs have to compute for weeks on end for the model to perform the way it does. 

The question of CAD 

While neural networks are great with ill-defined patterns in data they struggle with precision. Take for example the images DALL-E generates. From afar, they look sensible. If you zoom in, you’ll see that they're full of nonsensical things – people with six fingers, odd faces…It doesn’t get the figures exactly right. 

While this lack of precision isn’t always obvious in language, it’s a major hurdle in additive manufacturing, where CAD is frequently used. Machine learning models have difficulty working with CAD geometry because it’s very precisely defined. 

At Oqton, we use a lot of deep-learning models on geometry, and we’ve found that it's much easier to get them to work on meshes, voxels, or point clouds than on CAD. In a CAD model, you have a design tree. If you have the smallest error at the root note, you'll pay for it in the rest of the design tree. In a mesh, however, a mistake in one vertex won’t ruin the whole model. 

This is exactly why we’re investing heavily in making deep-learning models work on CAD. It’s difficult to get that precision, but that's where a lot of value lies. 

Automating repetitive tasks in manufacturing 

The main reason we are applying AI, specifically machine learning, to manufacturing is to capture a user’s knowledge, and leverage it to automate repetitive tasks.  

To get an idea of what manufacturing automation looks like, we need to look no further than the sheet metal industry. It’s incredibly automated, and it has been for the last 20 years. If you order a sheet metal part, you get a quote in five minutes. Orders come into the sheet metal shop, and the ERP will automatically generate nests and push them to the available machines. 

However, other industries – like additive, welding and machining – aren’t at that stage of automation yet.  

This is where the idea for Oqton came from. AI could be used to get the rest of the manufacturing space to the same level of automation as sheet metal, which would greatly increase productivity and facilitate innovation. 

As things stand, organizations are struggling to meet market demand. Batch sizes are getting smaller, there’s more interest in customization and personalized products, and new product introduction time is decreasing. Every new batch means that an engineer or technician needs to prepare data for printing anew or to adjust the workflow.  

It’s very difficult to find skilled labour that knows how to program advanced machines for five-axis milling and metal 3D printing. At the same time, existing staff is often stuck doing repetitive manual work. By automating the repetitive part of their work, staff could focus on solving problems instead.  

A verticalized approach to advancing manufacturing 

Oqton’s approach to manufacturing automatation was to create a single MES, IOT platform – the Manufacturing OS – that captures the complete digital thread of all the parts that are being produced. This information becomes the data engine the AI uses for automation. 

We first focused on metal and polymer 3D printing, and we decided to tackle the problem vertical by vertical. Instead of developing a single software for all industries, we created the best automation capabilities for a specific segment. 

Each vertical has similar types of geometries for which we can then train an AI, and apply it to all the organizations in that area. We started with dental labs, and our focus allowed us to train multiple machine learning models for crowns, bridges, RPD frames, dental models and clear aligners. 

Now we’re using the same approach for the healthcare industry, service bureaus, and the energy industry. 

AI-based automation is already here 

Several large organizations are already using the Manufacturing OS as the central coordinator.  Orders come in and Oqton manages the design, 3D printing, laser marking, CNC machining…the whole production. 

For smaller dental labs, Oqton provides the peace of mind that they won’t miss order deadlines due to staff shortages. Some of them have just one employee who knows how the printers, the programs and the processes work. If the employee is unavailable, printing grinds to a halt. However, Oqton’s AI models can capture that knowledge and enable the lab to run the machines at any time.  

This also allows more organizations to use additive technology. While before only a specialist could operate a machine, now the generalists in the organizations are empowered to use advanced technologies. 

As an added benefit, the AI helps improve utilization of the equipment by combining nesting and scheduling taking into account machine and operator availability.  

Crown Ceram, a dental lab in France, is one of our earliest customers. Their CEO, Frederic Rapp, recently shared his view on how AI from Oqton has helped his company, which nicely summarises what I’ve referenced above: 

“Oqton changed our workflow dramatically. The file preparation is AI-driven and therefore gets quicker and quicker. Right from the start we divided preparation time by two compared to the old method. 

“Oqton gave us the opportunity to boost our productivity through AI. We now can train our staff quicker, put more parts on the platform without compromising the end-result, and manufacture more parts per day with the same machines and quicker.” 

Reshaping manufacturing 

Looking back at the progress of AI solutions over the last decade and the results we’ve achieved in the dental industry, I am convinced that AI is becoming the defining technology of our age.  

This has been a long time in the making, and there have been a few false starts, but we’re there now. The level of AI capabilities that we can now build reliably will totally transform every part of our economy. 

The manufacturing industry is at the cusp of the next big wave of automation, where AI plays a vital role. In the same way that the PC transformed how office environments and factories were managed, AI is transforming production workflows and accelerating innovation. 

Looking for an in-depth look into how we use AI? Read our blog about AI-driven automation.