Introduction
Garbage patches. Landfills. Ocean waste. The average global citizen produces 2.6 lbs of trash a day, with Americans leading at 4.4 lbs/day. This waste is thrown into cities, dumps, and for some, recycling plants. But while we continuously fill up our recycling bins, only 9% of what we throw away actually gets recycled.
There are many different methods of recycling, from mechanical separation to thermal recycling, but one stuck out to us: biorecycling. Because so little gets broken down and reused, we need to find a way to break material like plastic down into its initial form - the monomers that make up your plastic bottles and caps and wrappers. We are unable to do that quickly and efficiently because of the contamination plastic faces every day, from stains on your paper pizza box to leftover juice in your solo cup. But with biorecycling, bacteria are able to break down all sorts of plastic in a relatively short period of time with little intensive care.
We have built two interconnected tools to aid in the biorecycling process. The first tool is a web app that recommends what bacteria combination and required conditions (e.g. temperature, pH, UV pre-treatment) are needed to breakdown plastic. The second tool is built on Google’s Teachable Machine, where scientists can upload files and verify what plastic they are breaking down. This integrates with our web app as it can be used to identify what plastic needs to be broken down and the web app provides recommendations for how to break them down.
Purpose and Motivation
The main question we have is this: why hasn’t anyone tried to do this before? Biorecycling has historically been extremely slow and studies have been done in very small amounts. There’s a lot that goes into this but namely, the biggest reasons for late adoption are a) it’s hard to integrate a new system into a process that’s already engrained (mechanical recycling or chemical plants) and b) it’s hard to test multiple rounds of experiments when some combinations take months to show any signs of progress. The iteration process is extremely slow.
Another reason for integration troubles is the fact that it’s hard to tell which plastic is which in a pile. Plastics have labels on them, anywhere from P1 to P8, but when they’re shredded or contaminated, it makes it extremely hard to tell which is which, and that slows down the process.
Our goal with this hackathon was to improve upon all three of these reasons. We wanted the system to be something scientists could use to test various bacterial combinations, an application recycling plants could use to verify which bacteria could be used to most efficiently break down various plastics, and an algorithm to understand which plastics are which and base those bacterial combinations off of it. There are constant strides in the biorecycling field every day, with more and more scientists coming up with ways to engineer bacteria to break down plastic quicker. If we can harness the power of AI and modeling tools to make plastic identification and bacterial pairing easier, we can aid in these new advancements and reduce the amount of plastic lost in the process.
How does it work?
Users choose what plastic they would like to breakdown by clicking on the button with the name of the plastic. Our first version includes Polyethylene (PET), Polyethylene terephthalate (PE), Polypropylene (PP), and Polylactic acid (PLA). When they clicked a buttons opens a page with a recommendation of what bacteria combination and required conditions opens. Under the recommendation there are three buttons that can be used to navigate to other recommendations for the plastic type. The “next option” button shows another recommendation, the “last option” button shows the recommendation before the current one, and the “all options” button shows all possible recommendations. To navigate to the home page, click “microbe match” in the left corner of the screen.
We would like to integrate a way to generate recommendations from an image of what plastic needs to be broken down. In our MVP, users can test out an early version of this feature by clicking on our Teachable Machine on the home page. We also provide a folder of sample files to use with the model. The Teachable Machine model identifies what type of plastic is present in one of the sample files, as a first step towards a computer vision model that identifies what plastic is needs to be broken down.
The Development Process
First, we had to understand the types of plastics, what items typically fell into those categories, and what bacteria could break them down. There are thousands of possible microbe combinations that could degrade a certain type of plastic. So we first began with searching research papers for verified microbe combinations that have synergistic degradation effects. However, to efficiently narrow down these combinations into which are most likely to degrade a certain type of plastic or combination of plastics, we needed to understand all the possible enzyme family groups including their traits, survival conditions, chemical groups and bonds. Through in-depth research, we found that enzymes that fall under the Oxidoreductase enzymes(Oxidases, Peroxidases, and Dehydrogenases) paired with each other are particularly potent to plastics containing aromatic compounds such as PET, PS, and PC. While Esterases such as the cutinases, lipases, and the recently famous FASTPETase, as well as many Proteases that are some of the best hydrolysis catalysts, are best for degrading plastics with ester bonds, such as PLA, that are susceptible to hydrolysis. Thus, by matching up microbes with similar living conditions, and their enzymes with synergistic relationships, thousands of accurate combinations can be produced, combinations that scientists can experiment with and further accelerate the concept of biorecycling.
Once we had this information, it was on to training. Our goal was to create models of items made of various plastics to train our Teachable Machine (TM) algorithm on. We used Fusion360 to model the items and applied a “physical material” onto the CAD models for each kind of plastic. We created two separate items for each of the four plastics we were sorting (medicine bottles and plastic caps for polypropylene - PP, plastic bottles and detergent bottles for polyethylene terephthalate - PET, plastic bags and plastic milk cartons for polyethylene - PE, and yogurt containers and solo cups for polylactic acid - PLA). Then, we recorded our screen, moving the model side to side from all different angles. We were able to cut the video down into separate screens to then feed into the TM model and after training, we could upload an image of both the CAD file and real-life items and tell which plastic they were made of.
We created the first version of our web app with HTML, CSS, and Javascript. Replit was used as a development environment as its easy to test while building and is collaborative. We updated our code on Github while iterating. Our Github is connected to Netlify where the web app is hosted. When a change is committed to Github it automatically updates to our site.
How to Use MicrobeMatch!
Head on over to microbe match! When you go to the website, you will see four different buttons labeled with different kinds of plastics and two links: one for a Google Drive with preloaded images and another with our Teachable Machine algorithm. Click the Google Drive link. You will be directed to two folders. Click the one titled “PLASTIC ASSETS” and download the images. Once you do that, redirect over to Teachable Machine and start testing the algorithm!
When you upload a file (not applicable to webcam), the algorithm will tell the user which type of plastic the item is. Feel free to test it out with any of the images in the folder (and try some from the internet as well, just for fun!) The goal of this step is to select an item commonly mis-recycled and identify which type of plastic it is. Once you’ve achieved this step, click back onto microbe match and select the plastic it belongs to. Once you do, it will showcase different kinds of bacterial combinations and conditions for optimal growth. You can showcase all the current options, click on sources, or view them all one by one.
Difficulties and Challenges
The recycling process is quite complex, with mechanical recycling involving 10+ different machines and chemical recycling only being possible in certain conditions. We spent a lot of time in the research phase to understand the different facets of these environments, but that made it difficult time-wise to finish our MVP and in the future, we plan to have more of an iterative phase of our process that isn’t as scrunched together.
With our application surrounding plastic identification, it took a long time to train the Teachable Machine (TM) model on both the individual CAD diagrams and regular objects. First off, we had to record a screen with the CAD model inside it and spin it at all different angles to collect all the data. From there, we had to split that video into separate snapshots of all the different angles and upload it to TM. This gave us a LOT of data. At one point, we had over 10,000 images loaded into the program which, in the end, actually ended up confusing the algorithm even more. We had to retrain it with fewer images that were more carefully selected to produce the correct results. CAD modelling was also a difficult task since we had to distinguish lots of different plastic objects from each other, amounting to 16 tests with three images each that all took around 5 minutes each time to upload and update. Eventually, we got the plastic identification to work, and we’d like to continue updating the prototype with images other than the ones listed in our Google Drive folder to better enhance the model.
In addition to difficulties with the plastic identification software, there were no existing lists, charts, or other online resources that outlined all of the possible enzymes (and microbes they belonged to) nor combinations they could form to degrade plastics. Databases that had some information were limited in that they only displayed microbes from studies that had been replicated several times, and did not provide possible untested combinations that scientists could experiment with. The biggest challenge is that it is not possible to know how many enzymes exist, just like we cannot know how many species could have ever existed on Earth. So, we began by researching the most well-known enzyme family groups that degrade plastics, then branched off into any sub-groups we learned throughout this research. Eventually, we created a map outlining all the major known enzyme family groups, the plastics they degrade, and examples of microbes that secrete those enzymes.
A challenge we faced when developing the web app was connecting the Teachable Machine to it. Originally we wanted to have an option to upload an image to the Teachable Machine on our site and a recommendation would be generated based off what plastic is detected. We did several tests and were able to connect models to our site using Tensorflow.js. However, it only had the option to use the webcam and not to upload a file. We decided to add a link to the Teachable Machine and the sample files to our home page for those who want to test out the first version. We still would like to connect it to our site and are going to do further research into how to do so.
Accomplishments and What We're Proud Of!
Valkyrie: I’m proud of how we used CAD modelling to illustrate different recyclable materials and learned Teachable Machine! I’m not too used to coding and AI and it was a pretty new experience for me, but it was also super fun once I got the hang of it trying to figure out how to train most effectively. I used to be super into CAD modelling a couple years ago and I’m also proud of the fact that I could pick it back up, relearn some of the skills, and put it to use for a project in sustainability!
Julia: I’m proud of how we were able to organize plastic-degrading enzymes and display them into a way that could help researchers make more incredible breakthroughs in bio recycling beyond FASTPETase. I love to work with biological materials and found it fun to take on a huge challenge like organizing enzymes and taking my expertise to the next level from high school biology to something more abstract that nobody really knows about and has a lot of “maybes” in.
Adara: I'm proud of what my team has accomplished during this hackathon as it was our first time exploring chemical and biological recycling. I'm proud of making the web app, as I've been learning how to code this summer. I wanted to participate in this hackathon as I love working on environmental projects and wanted to challenge myself to learn coding through building. I feel I learned a lot through working on this challenge and would like to keep iterating from our MVP!
What's next for Microbe Match (GTM)
Our goal for this project was to make a tool that was able to plug into existing operations. In both chemical recycling and mechanical recycling, our tool fits as either a) an identification system for various types of plastic or b) an information system to speed up the process of verifying bacterial combinations. By synthesizing all this data, we give researchers in the lab more time to actually perform their experiments while also reducing the friction between sorting plastics for workers in the recycling plants.
Our go-to-market strategy starts with us reaching out to researchers to beta test, iterate quickly, and fail forward to really iron out the kinks in the model. It doesn’t always specify that a plastic is 100% PET or PLA so that would be something we have to focus on more closely. There would also have to be add-ons for other types of plastic (our MVP has four, there are 8+ kinds of plastic readily used), that we’d need to work with scientists to find a solution for.
We also plan to, in the beginning, make the application easily accessible online for people to test out and add data to. From there, we can get a ton of data and sort through for the best pieces to train that would make it more applicable in real-world scenarios (almost like crowdfunding for data analytics). Once all of these pieces are in place, we’d like to start charging both recycling plants and researchers by usage through a subscription or pay by recommendation.
Built With
- autodesk-fusion-360
- css
- google-teachable-machine
- html
- javascript
Log in or sign up for Devpost to join the conversation.