A programmable chemical computer with memory and pattern recognition

Current computer architectures are constrained by the von Neumann bottleneck, separating processing and memory. Chemical systems offer a potential solution as processing and memory coexist, utilizing chemical reactions for computation. However, their lack of programmability has been a major limitation. This research introduces a programmable chemical processor to address this challenge. Unlike conventional computers reliant on binary electronic switches, this approach leverages the parallelism inherent in chemical systems. The design draws inspiration from naturally occurring parallel processing systems and previous research exploring the computational capabilities of coupled oscillator systems, such as those based on chemical, chemo-mechanical, spin-torque, or laser-based oscillators. These systems, while demonstrating computational potential, often suffer from weak coupling and susceptibility to noise. The current work builds upon the Belousov-Zhabotinsky (BZ) reaction, which has previously been used to emulate logic gates, perform image processing, and solve optimization problems. This study aims to overcome the limitations of previous approaches by combining individually addressable BZ reaction cells with advanced image processing techniques to enable programmability and flexible, multi-purpose calculations.

The paper reviews existing unconventional computing architectures, focusing on coupled-oscillator systems. These systems, which have been implemented using various methods including chemical, chemo-mechanical, spin-torque, and laser-based oscillators, often suffer from weak coupling and noise. Reaction-diffusion systems, particularly the BZ reaction, have shown promise in demonstrating solutions to pathfinding problems, implementing logic gates, and performing image processing and pattern recognition. Previous research using BZ reactions focused on specific problem-solving platforms exploiting spatiotemporal oscillations for information processing. This study builds upon these advancements by introducing a programmable platform capable of more versatile calculations.

The research utilizes an automated BZ reaction platform consisting of four main components: a customizable BZ reactor cell grid; a magnetic stirrer array to control the BZ reaction in each cell; a control interface connected to a computer to control the stirrers; and a camera and image processing system for monitoring and analyzing the BZ reaction in real-time. The BZ reaction itself uses ferroin ([Fe(Bpy)₃]²⁺/³⁺) as a catalyst, with color changes (red/blue) indicating the reduced and oxidized states. The programmability is achieved by controlling the magnetic stirrers in each cell, enabling independent control of local BZ oscillations. The interaction between cells is managed through hydrodynamic coupling, where the stirring patterns in one cell influence the oscillations in neighboring cells. Different grid designs, including one with V-shaped openings and one without, were tested to optimize inter-cell fluid propagation and improve pattern coherence. Image processing, using an SVM trained on a manually labeled dataset, is used to classify each cell's state (on/off) from video recordings of the BZ reaction. For pattern recognition, a reservoir computing approach was employed, where the BZ reaction's output was fed into a neural network for classification. A convolutional neural network (CNN) in TensorFlow was trained on datasets created from processed video data, representing different input patterns. In addition, a digital encoder, implemented as an Autoencoder Neural Network, was used for comparison purposes. This allowed a direct comparison of the BZ platform's computational capabilities with a purely digital system.

The researchers successfully demonstrated a programmable chemical processor using the BZ reaction. The platform exhibits more than 2.9 × 10¹⁷ chemical states, allowing for complex information encoding. Programmability was achieved via individual cell control using magnetic stirrers, manipulating the BZ oscillations and their propagation through hydrodynamic coupling. Two different grid designs were tested to optimize the coupling of BZ cells, enhancing the platform's computational capabilities. The system was shown to reliably store and process information, acting as a type of volatile memory due to the transient nature of BZ oscillations. The platform's ability to recognize patterns was evaluated using a reservoir computing scheme. The experiments revealed that the system is able to distinguish 20 different patterns with 92.5% accuracy using a CNN trained with data obtained through image processing of the BZ reaction. This accuracy was achieved using a dataset based on 20-time-point windows of data, resulting in 500 features per input pattern. The comparison to a digital autoencoder demonstrated that the chemical processor performs similar computational tasks, providing a benchmark for its performance. The researchers estimated that their BZ platform, based on a 1-minute window of data, encodes similarly to a digital encoder that would require 60 million logic gates, enabling the decoding of the equivalent of 1 million logic gates per oscillation. This highlights the platform's effectiveness as a chemical encoder.

This research successfully demonstrates a programmable chemical computer that transcends the von Neumann bottleneck. The system's high-dimensional state space enables complex pattern recognition with considerable accuracy, exceeding the capabilities of previously reported chemical computing approaches. The use of the BZ reaction, coupled with the precise control offered by magnetic stirrers and advanced image processing, enabled the creation of a fully programmable platform. The findings suggest that chemical computing using reaction-diffusion systems represents a viable alternative to traditional silicon-based computation. The comparison to a digital autoencoder provides a compelling argument for the platform's capacity to perform complex information processing tasks, showcasing its potential as a chemical computing paradigm.

This work successfully demonstrated a programmable chemical computer based on the BZ reaction, showing that a large number of states can be used to encode and decode information and that this chemical processor can recognize patterns with a relatively low error rate. While limitations exist concerning reagent depletion and oscillation rate, the system's potential for high parallelism and low power consumption suggests that it could provide a significant advance in computational paradigms beyond the von Neumann architecture. Future research could focus on miniaturizing the BZ cells to allow for spontaneous cell-cell interactions, thus reducing the reliance on electronic control, and on further refining the pattern recognition algorithms to reduce reliance on machine learning.

The study notes that the BZ reaction reagents deplete over time, limiting the duration of continuous computation. The oscillation rate of the BZ reaction, while tunable, is currently slower than that of digital computers. The current reliance on a camera for read-out and machine learning for pattern recognition represents a hybrid system, and further research is needed to reduce these external dependencies, ultimately aiming toward a fully autonomous chemical computer.