What happens when you apply machine learning to 3D printing?

28 Jun 2021

Dr Vivek Mahato. Image: I-Form

Researching the future of manufacturing at I-Form, Dr Vivek Mahato believes intelligent systems can help to further reduce waste and improve product quality.

After achieving a bachelor’s degree, Dr Vivek Mahato started his career as a web developer in India. But his interest in researching data science drew him to Ireland.

He completed a master’s degree in University College Dublin, followed by a PhD in machine learning in additive manufacturing (more commonly known as 3D printing) at I-Form, the Science Foundation Ireland research centre for advanced manufacturing.

While UCD is the lead institution for I-Form, its research spans other academic partners, including Dublin City University (DCU). It’s here that Mahato took up a postdoctoral position in order to extend the research he began under his PhD.

‘Curiosity, along with an aptitude for coding, propels my research career in machine learning’

What inspired you to become a researcher?

I see myself back in fifth grade attending a class of Mr Sanjay Adhikary, our science teacher. His teaching excellence and captivating stories drove me to be curious.

As I grew up, I delved more into finding new avenues to perform a task, other than taught traditional methods. Sometimes the approach was more complex, and sometimes I found a more straightforward way.

Curiosity, along with an aptitude for coding, propels my research career in machine learning, and I relish it.

What research are you currently working on?

The accessibility of high computation power and in-situ additive manufacturing (AM) process data has promoted more complex data-driven mechanisms. With time, the research questions have evolved from analysing the macro properties of a product to predicting its micro properties.

In our research, we try to predict the porosity of the product while it is being manufactured. Our machine learning (ML) assessment runs parallel to the AM process in real-time. Our models perform efficiently in the task of classifying individual raster scans as normal or anomalous. This study was conducted at I-Form under the supervision of Prof Pádraig Cunningham (UCD) and Prof Dermot Brabazon (DCU).

As a next step, we are researching making ML models more robust and process agnostic. Our ongoing research, with the help of Dr Annalina Caputo (DCU), also targets parameter recommendation and optimisation of an AM process.

In your opinion, why is your research important?

Having predictive control over an additive manufacturing process has direct benefits and impact. During manufacturing, if any anomaly or a defect which could impact the quality of the product arises (like a pore), the system could flag the incident and notify the operator. The operator can then tune the parameter, using the recommended settings, to mitigate future incidents and re-melt the laid layer to remove the pores that occurred.

Such an intelligent system helps reduce material wastage and energy cost to a greater degree. It also helps to govern the product quality and the throughput of the manufacturing plant.

ML assessment of the product’s quality also promotes less reliance on expensive and slow measurement techniques to gauge the quality of the part after production.

What commercial applications do you foresee for your research?

Despite advancements in metal AM technology, the difficulties for obtaining optimal process control and process repeatability are still significant. We can tackle this challenge by richly studying the AM process and employing data-driven approaches and tools. Therefore, we can build a more robust and efficient process that reduces waste of materials and energy.

The potential to apply these technologies in a commercial setting is massive. Traditional manufacturing is a subtractive process, where a big block of material is sliced to make a part. Therefore, the method limits product design flexibility and generates a lot of wastage as leftover materials.

Using an intelligent AM system, we would be manufacturing products customised to the application’s needs to a greater extent, and reducing the overhead production and quality-check costs, as well as environmental impact.

What are some of the biggest challenges you face as an additive manufacturing researcher?

In my opinion, AM is relatively new on the global platform. The major challenge is the limited access to informative open-source production data in the research community. Therefore, access to production data is restricted to the ones generated from in-house laboratories.

However, as the research in this domain is rising, I believe that the availability of the production data will not be a hurdle anymore.

Besides, with AM, significant up-front expenditure in tools can be expected and some specialised powders are costlier than common raw materials.

Are there any common misconceptions about additive manufacturing?

When one hears ‘additive manufacturing’ or ‘3D printing’, the common misconceptions often fall into: ‘3D printing is printing stuff made of plastics’; ‘AM manufactured products are not that strong’; or ‘AM is expensive and difficult to operate’.

Though 3D printers in the past were quite complex to operate, 3D printers now come with user-friendly software that does not require much expertise. There is an abundance of 3D printers for under €250 that use thermoplastics to construct a product.

As we progress and develop the field of AM, the systems available on the market are getting more accessible, and budgets can stretch to specialised printers. A 3D printer now has the flexibility of choosing among a wide array of raw materials like glass, metals (steel, titanium or gold) and bio-ink created using stem cells.

We can tune the AM process to customise the part’s properties – for example, tensile strength, rigidity and thickness. Therefore, depending upon the application, manufacturing strong and solid parts is possible.

What are some of the areas of research you’d like to see tackled in the years ahead?

Traditional physical and statistical models effectively simulate the manufacturing process. However, they are pretty slow and precise to a specific task or production.

To be accurate, a deep understanding of the materials and the build chamber’s environment is required. Machine learning assessment does not address this challenge as it tries to discover patterns directly from the data. There is a potential of utilising these approaches in combination, to simulate and validate a manufacturing process.

There is also a need to enhance the repeatability of an AM process to keep product quality consistent.

Are you a researcher with an interesting project to share? Let us know by emailing editorial@siliconrepublic.com with the subject line ‘Science Uncovered’.