Skip to main content

Brain-Inspired Organic Modular Lab-on-a-Chip for Cell Classification

Periodic Reporting for period 2 - BIOMORPHIC (Brain-Inspired Organic Modular Lab-on-a-Chip for Cell Classification)

Reporting period: 2020-07-01 to 2021-12-31

Software systems known as neural networks are at the heart of artificial intelligence and are becoming more and more part of our daily life. These networks are used to translate languages, classify images, and recognize diseases, among many other applications. Deep-learning algorithms require a substantial amount of computer resources, but can nevertheless be executed by conventional computers, built around the classical von Neumann architecture. However, the simulation of these networks on classical systems turns out to be extremely inefficient. Brain-inspired (neuromorphic) computing aims to mimic the efficiency of the brain in such A.I. applications and has recently demonstrated advancements in pattern and image recognition. However, conventional neuromorphic devices and systems suffer from instability, unpredictability, and often substantial energy use, complicating the path to achieve the interconnectivity and efficiency of the brain.

In this project we use a low-cost solution organic artificial synapse as a building-block for these smart systems. The conductance of this single synapse can be accurately tuned by controlled ion injection in the conductive polymer which allows for straightforward low-energy analogue computing. BIOMORPHIC is building an interconnected network of these synapses to obtain a true neuromorphic array. The goal is to develop a unique brain-inspired organic lab-on-a-chip in which microfluidics integrated with sensors, collecting characteristics of biological cells, will serve as input to the neuromorphic array. BIOMORPHIC will combine modular microfluidics and machine-learning to develop a novel platform for low-cost lab-on-a-chip devices capable of on-chip classification. Ultimately the project aims to classify circulating tumour cells on the chip using a multitude of sensor signals trained to recognize these cells.
The first steps include materials research and optimization, and the development of a stand-alone neuromorphic array that can be trained on-chip and fully in hardware. We performed experiments to investigate the stability of the neuromorphic devices on the chip. Particularly important is the state retention of the individual devices. This state retention is important to be sufficient for the duration of the training and classification of data. As a demonstration of the on-chip learning we have developed a modular biosensor that is trained to classify a model disease from sweat samples. The concept behind this approach is that the ion sensors have a straightforward output signal that can be used as inputs to the hardware neural network for classification. Since these signals can be linearly separated, this smart biosensor can be trained on the chip, with the need for complicated error backpropagation with multiple layers. This first smart biosensor chip includes a small amount of devices that can be individually addressed. We are still investigating the large-scale arrays to upscale to a multitude of devices using crossbar architectures.

Simultaneously, the microfluidic chip including its sensors is being developed. We have designed, developed, and fabricated an impedance sensor as well as an optical sensor integrated within a microfluidic chip and used these sensors to differentiate between different microbeads as well as some first type of cancer cells.

Finally, we have successfully developed new (polymer) materials that run in enhancement-mode and thus have larger dynamic ranges (ON/OFF ratios) and operate with lower power. At the same time, we have integrated a solid electrolyte so that the on-chip biosensor neuromorphic hardware can run without fear of water contamination of electronics.
Neuromorphic computing uses hardware to efficiently perform computational demanding tasks in artificial intelligence. However, the training of these networks still relies on external computers that write to the hardware. In our research we are the first to perform on-chip learning without the help of external computers, fully in hardware. These results are certainly beyond the state of the art. In fact, as we use materials that can be easily trained on-chip, we can also use them to retrain using different input signals, highlighting the versatility of this platform by repurposing the same chip for the classification of other data. This is also beyond the state of the art and has not been shown before. Finally, for classification of cancer cells, it is common to rely on one particular sensor, where we are targeting the combination of multiple sensors. We expect these capabilities to result in a diverse and versatile platform that can be used to classify cells, or other information from sensors, locally and on the chip. However, the retraining and repurposing of this hardware neural network allows for a wider application field, and will demonstrate the capabilities of local edge-computing devices that can be trained and retrained with new data.