Periodic Reporting for period 2 - SHAPE (Synaptic Switching with Halide Perovskites)
Okres sprawozdawczy: 2022-07-01 do 2023-12-31
Our daily lives are increasingly dominated by the digital world. Ubiquitous access to the internet, the rise of artificial intelligence (AI) with increasingly practical uses, and the connection of all devices around is means that the need for computational power is constantly increasing. As a result, large server farms now consume a significant faction of our energy, and this fraction is set to increase further in the coming decades.
Many of the energy-intensive tasks use neural networks, including immensely popular applications like generative AI, as is used by text generating engines like ChatGPT and image generators like Midjourney. Neural networks consist of nodes in a so-called layer, that are connected to another layer of nodes. The complexity and power of a neural network increases with an increasing number of layers and nodes. The connections are the key to the working mechanism of the neural network. They can vary in strength, which means that a strong connection transmits more of the signal than a weak connection. These are analogous to the biological synapses in the human brain, while the nodes take on the function of the neuron.
In artificial neural networks, we typically simulate the components (notes and connections) using traditional transistors. To simulate one component one needs to employ several (tens) of transistors, which leads to a very high the energy consumption of a neural network when compared to a biological brain. In this project we aim to bring the energy consumption of artificial neural networks closer to the energy consumption of the human brain by inventing devices that emulate the connections/synapses using artificial semiconductors, in particular metal-halide perovskites. These semiconductors are unique because they conduct both ions and electrons very efficiently. We use the fast electron motion to read and write the synapse, and the much slower ionic motion to memorize the conductive state of the synapse, which is the strength of the connection.
To achieve these goals the program is divided in two parts. In one part we develop a better understanding of ion migration in metal halide perovskites. In the second part we then use this understanding to fabricate artificial synapses with very low energy consumption, approaching values close to those of the human brain.
Many of the energy-intensive tasks use neural networks, including immensely popular applications like generative AI, as is used by text generating engines like ChatGPT and image generators like Midjourney. Neural networks consist of nodes in a so-called layer, that are connected to another layer of nodes. The complexity and power of a neural network increases with an increasing number of layers and nodes. The connections are the key to the working mechanism of the neural network. They can vary in strength, which means that a strong connection transmits more of the signal than a weak connection. These are analogous to the biological synapses in the human brain, while the nodes take on the function of the neuron.
In artificial neural networks, we typically simulate the components (notes and connections) using traditional transistors. To simulate one component one needs to employ several (tens) of transistors, which leads to a very high the energy consumption of a neural network when compared to a biological brain. In this project we aim to bring the energy consumption of artificial neural networks closer to the energy consumption of the human brain by inventing devices that emulate the connections/synapses using artificial semiconductors, in particular metal-halide perovskites. These semiconductors are unique because they conduct both ions and electrons very efficiently. We use the fast electron motion to read and write the synapse, and the much slower ionic motion to memorize the conductive state of the synapse, which is the strength of the connection.
To achieve these goals the program is divided in two parts. In one part we develop a better understanding of ion migration in metal halide perovskites. In the second part we then use this understanding to fabricate artificial synapses with very low energy consumption, approaching values close to those of the human brain.
performed measurements and simulations towards this aim. The measurements of ion migration in different perovskite solar cells were performed using the capacitance of these solar cells over time after applying a voltage pulse. The voltage pulse drives ions through the device, which changes the capacitance. When releasing the voltage pulse, the ions move back, and from the change of capacitance we can learn about the properties of the ions.
In the meanwhile we noticed that the simple model applied to these capacitance measurements is not valid for metal halide perovskites. We therefore perform drift-diffusion simulations of these devices to indicate which quantitative information can be extracted.
In the second part of the project, we aim to develop an artificial synapse from metal halide perovskites with a very low energy consumption. This aim should be achievable, because the migration of ions in these perovskites takes very little energy. Towards this aim we have downscaled the perovskite device and now achieve an energy consumption to switch a memristive state that is lower than one picojoule. At such low energy consumption, the capacitance of the measurement system can be larger than the device capacitance and thereby overshadow the measurement. We are developing an new measurement setup with very low capacitance to measure an energy consumption in the femtojoule regime.
In the meanwhile we noticed that the simple model applied to these capacitance measurements is not valid for metal halide perovskites. We therefore perform drift-diffusion simulations of these devices to indicate which quantitative information can be extracted.
In the second part of the project, we aim to develop an artificial synapse from metal halide perovskites with a very low energy consumption. This aim should be achievable, because the migration of ions in these perovskites takes very little energy. Towards this aim we have downscaled the perovskite device and now achieve an energy consumption to switch a memristive state that is lower than one picojoule. At such low energy consumption, the capacitance of the measurement system can be larger than the device capacitance and thereby overshadow the measurement. We are developing an new measurement setup with very low capacitance to measure an energy consumption in the femtojoule regime.
The project has so far contributed to a better understanding of ion migration in perovskites, and we have developed an artificial synapse with an energy consumption that is lower than any previous artificial synapse from halide perovskites. In the next reporting period, we will develop a new analytical model to better understand the capacitance transient measurements we are performing on perovskite devices. With the new model, we can use these capacitance measurements for a quantitative understanding of ion migration in perovskites. We will further downscale the artificial synapses to lower the energy consumption to about 100 femtojoule. At this point the energy consumption is as low as the energy consumption of biological synapses. Finally, we will work on understanding of the working mechanism of artificial synapses made from halide perovskites.