Business Problem
- The grade of plastic manufactured is a combination of input raw material quality & quantity, which decides the tensile strength & other characteristics, making it suitable for manufacturing specific types of plastic products
- The polymers sourcing market, being highly un-managed, has a host of supply chain issues & pricing fluctuation, leading to the unpredictability of the end product costs & quality
- To overcome this, Atom Corporation had set up an R&D unit of 25 people to manually test different combinations of these raw materials to produce the end product in multiple, controlled ways
- The manual work required more than 30 days to produce the correct raw material manufacturing composition. Therefore, much wastage increased the order fulfilment time by 50%, resulting in order pile-up & customer loss.
- GoML worked helped build an ML-based solution to predict the correct composition of raw polymers balls to produce a grade of plastic with certain characteristics
Solution
- We worked with the team at Atom Corp. to understand the data that R&D team churns & the overall experimentation process
- Manually collected data from the last 7 years was shared with us in excel files, which was loaded into S3 bucket
- The data was then cleaned, removing the data where records of experiments which were done in non-standard environments & conditions and the Preprocessed data was stored in RDS SQL Database by leveraging SQL jobs
- Amazon SageMaker was leveraged to experiment (Python , Pandas, Scikit learn) with 3 models Mixed logistic regression, Decision tree classification & SVM classification
- It was found that a certain configuration of Decision tree classification performed as per expectation giving out an accuracy of 89%. The model was then deployed with production data on Amazon SageMaker
Architecture
Tools Used
Outcomes
