A fully-connected annealer extendable to a multi-chip system and featuring a multi-policy mechanism has been designed by Tokyo Tech researchers to solve a broad class of combinatorial optimization (CO) problems relevant to real-world scenarios quickly and efficiently. Named Amorphica, the annealer has the ability to fine-tune parameters according to a specific target CO problem and has potential applications in logistics, finance, machine learning, and so on.

The modern world has grown accustomed to the efficient delivery of goods right at our doorsteps. But did you know that realizing such efficiency requires solving a mathematical problem, namely what is the best possible route between all the destinations? Known as the “traveling salesman problem,” this belongs to a class of mathematical problems known as “combinatorial optimization” (CO) problems.

As the number of destinations increases, the number of possible routes grows exponentially, and a brute force method based on an exhaustive search for the best route becomes impractical. Instead, an approach called “annealing computation” is adopted to find the best route quickly without an exhaustive search. Yet, a numerical study done by Tokyo Tech researchers has shown that while there exist many annealing computation methods, there is no one method suitable for solving a broad class of CO problems. Therefore, there is a need for an annealing mechanism that features multiple annealing methods (a multi-policy mechanism) to target a variety of such problems.

Fortunately, the same team of researchers, led by Assistant Professor Kazushi Kawamura and Professor Masato Motomura from the Tokyo Institute of Technology (Tokyo Tech), have reported a new annealer that features such a multi-policy approach or “metamorphic annealing.” Their findings are published in the Proceeding of ISSCC2023 and will be presented in the upcoming 2023 International Solid-State Circuits Conference.

“In the annealing computation, a CO problem is represented as an energy function in terms of (pseudo) spin vectors. We start from an initially randomized spin vector configuration and then update it stochastically to find the minimum energy states by reducing its (pseudo) temperature. This closely mirrors the annealing process of metals where hot metals are cooled down in a controlled manner,” explains Dr. Kawamura. “Our annealer named Amorphica features multiple annealing methods, including a new one proposed by our team. This provides it the ability to adopt the annealing method to the specific CO problem at hand.”

The team designed Amorphica to address the limitations of previous annealers, namely that their applicability is limited to only a few CO problems. This is first because these annealers are local-connection ones, meaning they can only deal with spin models having local inter-spin coupling. Another reason is that they do not have flexibility in terms of annealing methods and parameter control. These issues were solved in Amorphica by employing a full-connection spin model and incorporating finely controllable annealing methods and parameters. In addition, the team introduced a new annealing policy called “ratio-controlled parallel annealing” to improve the convergence speed and stability of existing annealing methods.

Additionally, Amorphica can be extended to a multi-chip, full-connection system with reduced inter-chip data transfer. On testing Amorphica against a GPU, the researchers found that it was up to 58 times faster while using only (1/500) power consumption, meaning it achieves around 30k times more energy efficiency.

“With a full-connection annealer like Amorphica, we can now deal with arbitrary topologies and densities of inter-spin couplings, even when they are irregular. This, in turn, would allow us to solve real-world CO problems such as those related to logistics, finance, and machine learning,” concludes Prof. Motomura.

A study by an international team of scientists shows that an irreversible loss of the West Antarctic and Greenland ice sheets, and a corresponding rapid acceleration of sea level rise, may be imminent if global temperature change cannot be stabilized below 1.8°C, relative to the preindustrial levels.

Coastal populations worldwide are already bracing for rising seas. However, planning for counter-measures to prevent inundation and other damages has been extremely difficult since the latest climate model projections presented in the 6^th assessment report of the Intergovernmental Panel on Climate Change (IPCC) do not agree on how quickly the major ice sheets will respond to global warming.

Melting ice sheets are potentially the most significant contributor to sea level change, and historically the hardest to predict because the physics governing their behavior is notoriously complex. “Moreover, computer models that simulate the dynamics of the ice sheets in Greenland and Antarctica often do not account for the fact that ice sheet melting will affect ocean processes, which, in turn, can feed back onto the ice sheet and the atmosphere,” says Jun Young Park, a Ph.D. student at the IBS Center for Climate Physics and Pusan National University, Busan, South Korea and first author of the study.

Using a new supercomputer model, which captures the coupling between ice sheets, icebergs, ocean, and atmosphere for the first time, climate researchers found that an ice sheet/sea level run-away effect can be prevented only if the world reaches net zero carbon emissions before 2060.

“If we miss this emission goal, the ice sheets will disintegrate and melt at an accelerated pace, according to our calculations. If we don’t take any action, retreating ice sheets will continue to increase sea level by at least 100 cm within the next 130 years. This would be on top of other contributions, such as the thermal expansion of ocean water” says Prof. Axel Timmermann, co-author of the study and Director of the IBS Center for Climate Physics.

Ice sheets respond to atmospheric and oceanic warming in delayed and often unpredictable ways. Previously, scientists have highlighted the importance of subsurface ocean melting as a critical process, which can trigger runaway effects in the major marine-based ice sheets in Antarctica. “However, according to our supercomputer simulations, the effectiveness of these processes may have been overestimated in recent studies,” says Prof. June Yi Lee from the IBS Center for Climate Physics and Pusan National University and co-author of the study. “We see that sea ice and atmospheric circulation changes around Antarctica also play a crucial role in controlling the amount of ice sheet melting with repercussions for global sea level projections,” she adds.

The study highlights the need to develop more complex earth system models, which capture the different climate components and their interactions. Furthermore, new observational programs are needed to constrain the representation of physical processes in earth system models, particularly from highly active regions, such as Pine Island glaciers in Antarctica.

“One of the key challenges in simulating ice sheets is that even small-scale processes can play a crucial role in the large-scale response of an ice sheet and for the corresponding sea-level projections. Not only do we have to include the coupling of all components, as we did in our current study, but we also need to simulate the dynamics at the highest possible spatial resolution using some of the fastest supercomputers,” summarizes Axel Timmermann.