IBM Research: When AI and quantum merge

IBM’s research laboratory in Zurich.

Judith Linine – Shutterstock.com

Above Lake Zurich, in the hills of Rüschlikon, the next chapter in industrial history is currently being written. While the world discusses the latest advances in generative AI, Alessandro Curioni’s team at IBM Research – Zurich, IBM’s European research center, is already working on a fundamental transformation that goes far beyond mere software updates.

When visiting the research center, one can only guess that history may be in the making here. From the outside, IBM Research, which has been located in Rüschlikon since 1962, looks about as “attractive” as a German school from the 1970s. Only the numerous cameras hint at something special going on inside.

And indeed it is: the laboratory has produced four Nobel Prize winners to date. Researchers are currently working on nothing less than the dawn of a new era in which artificial intelligence (AI) and quantum computing merge synergistically. This should enable problems that were previously considered unsolvable to be solved using conventional methods.

According to CEO Curioni, we are currently experiencing two of the biggest technological changes of the last 30 years at the same time. For Curioni, the core of this change lies in a redefinition of computer science.

• AI is changing abstraction. Instead of laboriously translating problems into complex mathematical equations (such as the Navier-Stokes equations in fluid dynamics), foundation models allow direct abstraction from data. This can reduce computing times in industry from 30 minutes to ten seconds—with an accuracy of a few percent deviation.
• Quantum computing changes representation. This is no longer about the binary logic of 0 and 1 in the digital era. Quantum computers use the Bloch sphere to represent information in a continuous space. This makes it possible to directly map the structure of nature—for example, in chemistry or materials science—instead of merely simulating it inadequately.

Quantum-centric supercomputing

However, this change does not automatically mean a departure from classic CPUs and GPUs, even if IBM’s IT vision is “””quantum-centric supercomputing.” In this architecture, the quantum processor is no longer just an exotic accelerator, but the valuable core around which classic CPUs, GPUs, and AI models are orchestrated.

A look inside an IBM Quantum System Two.

Hill

The hybrid approach taken by the major bank HSBC provides a first impression of what is possible with this technology. In a pilot project on algorithmic trading with corporate bonds, a team at the bank used quantum computers for “feature engineering.” The result: a 34% increase in prediction accuracy compared to traditional models. Although Philip Intallur, Head of Quantum Technologies at HSBC, is reluctant to call this a quantum advantage in the academic sense, the clear commercial advantage is obvious to him.

However, as the performance of quantum computers grows, so does the threat to current encryption methods. The “harvest now, decrypt later” attack—stealing data today to decrypt it in ten years—is a real danger. IBM and partners such as HSBC are therefore pushing for a ten-year modernization phase toward quantum-secure cryptography standards (post-quantum cryptography).

Intensive research is therefore being conducted in Zurich laboratories on quantum-secure tapes for long-term archiving. These are essential if sensitive data in sectors such as government, military, finance, or healthcare is to be stored for periods of up to 30 years. And they must also be protected from decryption for 10-20 years. IBM researchers are currently considered leaders in the implementation of post-quantum cryptography in tape firmware.

Advances in tape development. On the left is a quantum-secure tape drive.

Hill

In parallel with its research into quantum computers, IBM is promoting sustainability through AI. One example of this is the ImpactMesh project. In collaboration with ESA and NASA, foundation models are being used for Earth observation to enable real-time disaster relief. In Kenya, these models have already helped to predict flood risks and landslides within days instead of weeks.

AI detects dangerous PFAS

AI is also finding its way into materials research. For example, the AI solution Safer Materials was developed to help companies identify toxic or environmentally harmful chemicals in their products and replace them with safer alternatives. One focus is on so-called forever chemicals (PFAS). AI can use structure-based analysis to determine whether a chemical belongs to the PFAS class, even if global definitions vary. It also suggests sustainable alternatives before regulations take effect.

IBM researchers in Zurich, in collaboration with Bane Nor, are demonstrating how AI models with a little intuition can be used for other purposes. The Norwegian state-owned company is responsible for the country’s railway infrastructure. This includes regularly checking 4,000 kilometers of track for defects. Until now, this has been a Sisyphean task, with inspectors often walking the tracks on foot to record damage to sleepers, rails, or fastenings.

AI on Norway’s railways

In collaboration with IBM Research, a process for automated visual inspection was developed. The core of the project is based on knowledge transfer. The IBM team used an AI model that was originally developed for inspecting concrete structures such as bridges. This was adapted to the Norwegian rail network.

Scanning tunneling microscope in one of the Zurich laboratories. The innovation won the Nobel Prize in Physics in 1986.

Hill

Without ever having seen rail data before, this model was immediately able to identify the main components of the infrastructure—such as sleepers and fasteners—in image data. The biggest hurdle to scaling was the sheer volume of data: measurement trains equipped with cameras deliver around 600,000 images per campaign. Since manual annotation of this enormous amount of data for AI training is impossible, the researchers relied on the method of visual prompting.

All that is required is for an expert to mark a single image of a defect (such as a broken sleeper fastening). The AI uses this visual prompt to automatically detect similar patterns in the remaining 600,000 images using feature matching technologies. This acts as a massive accelerator: instead of spending months preparing data, a precise model can be created in a very short time with minimal manual effort.

To further increase efficiency, various specialized AI models were merged into a single fusion model. This model can now simultaneously recognize components and locate specific defects, which significantly increases processing speed.

Computing power for AI

What began as a proof of concept (PoC) in 2023 has evolved into a scalable solution. It enabled Bane Nor to transition from rigid maintenance schedules to risk-based maintenance. Repairs are made where the AI sees the most urgent need for action. A mobile app was also developed for on-site employees, which visualizes the defects found directly on a map.

To handle such computing loads, IBM continues to rely on the mainframe. This includes, for example, the System Z17. Equipped with the Telum 2 processor and the Spyre AI Accelerator, it is designed to enable the analysis of financial transactions in real time. Another highlight from IBM’s perspective is the automated fight against money laundering (anti-money laundering): by combining graph-based machine learning and large language models (LLMs), it has been possible to increase the detection rate by 70 percent and massively reduce false alarms.