Project Overview
Plastics are used almost everywhere in our daily lives because they are cheap and light weight.
But using too much plastic and not recycling it properly is becoming a big environmental issue.
Traditional recycling methods like mechanical recycling do not work well for mixed or dirty
plastics. To solve this issue chemical recycling, especially pyrolysis, have become very popular
nowadays. Pyrolysis can turn plastic waste into useful products like fuels and chemicals.
However, it is hard to get the best results from the pyrolysis process because the reactions are
complex, the waste materials are different and many experiments are needed to convert the
waste plastic to some value added product. So, the use of Artificial Intelligence (AI) and
Machine Learning (ML) makes the process faster, more accurate and more efficient.
In the past few years the use of AI and sensing technologies in plastic waste management are
growing rapidly. Like, recently models are made for sorting the plastic and other materials so
that we can use the sorted plastic to further convert into more useful products. Further more,
models are made, for example, logistic regression (LR) for identifying different types of plastics
like polyethylene (PE), polyethylene terephthalate (PET), polypropylene (PP), polystyrene (PS),
and polyvinyl chloride (PVC) from the plastic waste (Fang et al., 2025). These technologies give
better solutions than old recycling methods (Joshi et al., 2025,Ahmed & Asadullah, 2020). Main
researches are being done in four areas related to this project are predicting yield, identifying
and sorting plastic types, monitoring the process using ML, and using robots for automation.
AI has also made sorting easier and faster. For example, Deep learning models like ResNet and
MobileNet, when used with near-infrared (NIR) spectroscopy, identifies different types of plastics
such as PET, HDPE, and PVC with up to 93% accuracy (Subedi et al., 2025).But these kind of
models are highly accuracy centric which is why they generally overlook real world challenges
like dataset biasness and label noise, due to which even a one incorrect decision may
contaminate whole material stream (Lubongo et al., 2024, Jeswani et al., 2021). Recently a
study was done on microplastic characterisation that even expert human writings may have
significant errors, emphasizing the need for reliability-aware AI systems (Herb et al., 2025).
This review critically analyzes the current status of AI-based plastic waste classification and
specifically targets the limitations of deterministic models and the less explored area of
uncertainty estimation. This paper identifies research gaps and provides future directions for
developing AI model for sorting plastic waste which tells about the uncertainty and
trustworthiness of the waste material.
Scope and Planning
Our initial phase focused heavily on market research, where we found that the plastic waste
sector is expected to hit around $42.71 billion by 2026. However, the real takeaway from our
research wasn't just the market size; it was the "purity gap." For chemical recycling like pyrolysis
to actually work, the industry needs over 96% purity in its feedstock. That’s an incredibly high
bar that current systems just aren't hitting.
The problem with current models is that they're built for "lab conditions." In the real world, you
deal with mislabeled data and literal dirt, which causes standard models to fail. We’ve defined
our project scope to tackle this head-on. Instead of just trying to get a higher accuracy
percentage, we are building a system that can "know when it’s wrong." By integrating
uncertainty estimation, our goal is to create a trustworthy sorting process where the AI flags
questionable materials for a human check rather than letting them contaminate the batch.
Currently, we’re refining our 1D-CNN to handle the specific overlapping signals found in NIR
spectroscopy data, which is where most of these purity errors start.
ML Implementation Plan
Phase 1: Data Strategy
● Input Data: Near-Infrared (NIR) spectra or Image data of PE, PET, PP, PS, and PVC.
● Pre-processing: Use Savitzky-Golay filtering (if using spectroscopy) to remove noise
and Standard Normal Variate (SNV) scaling.
● Uncertainty Labeling: Identify "borderline" samples that the model struggles with.
Phase 2: Model Architecture
● Backbone: Use a 1D-CNN (for spectra) or ResNet (for images).
● Uncertainty Mechanism: Add Dropout layers that remain active during testing. By
running the same sample through the model 10 times, you get a distribution of results. A
high variance in results = High Uncertainty.
We are implementing this using a Bayesian approach. Unlike standard models that provide a
single "best guess," our model performs multiple forward passes for every sample. By
calculating the standard deviation of these predictions, we generate an "Uncertainty Score."
● High Confidence: The material is automatically sorted.
● High Uncertainty: The material is flagged for manual inspection or secondary sensor
verification. This "human-in-the-loop" strategy ensures that the feedstock for chemical
recycling remains pure, directly addressing the industry's need for trustworthy
automation.
Code Framework
import torch
import [Link] as nn
import [Link] as optim
class TrustworthyPlasticCNN([Link]):
def __init__(self):
super(TrustworthyPlasticCNN, self).__init__()
self.conv1 = nn.Conv1d(1, 16, kernel_size=3)
self.fc1 = [Link](16 * 98, 64)
[Link] = [Link](p=0.5) # The key for uncertainty
self.fc2 = [Link](64, 5) # 5 Classes: PE, PET, PP, PS, PVC
def forward(self, x):
x = [Link](self.conv1(x))
x = [Link]([Link](0), -1)
x = [Link](self.fc1(x))
x = [Link](x) # Dropout stays ON during inference
return self.fc2(x)
# Function to estimate uncertainty
def predict_with_uncertainty(model, sample, iterations=10):
[Link]() # Keep dropout active
preds = [[Link](model(sample), dim=1) for _ in range(iterations)]
preds = [Link](preds)
mean_prediction = [Link](dim=0)
uncertainty_score = [Link](dim=0) # High std = Low trustworthiness
return mean_prediction, uncertainty_score