Breakthrough quantum technology slashes computational requirements for molecular modelling

In recent years, quantum computers have been garnering attention as a next-generation technology. From healthcare and materials development to the energy sector, their potential applications are virtually limitless. However, significant hurdles still stand in the way of practical implementation. In particular, the issue of qubit errors is severe, and achieving high-accuracy computations requires redundancy to correct these errors. Consequently, realistic computations would require around millions of qubits—a technology that currently remains far in the future.

To address this challenge, we have devised a groundbreaking approach. At Fujitsu Quantum*1, we have developed two core technologies: the “STAR Architecture ver. 3”, jointly developed by Fujitsu and The University of Osaka*2 based on the Fujitsu Small Research Lab*3 and “molecular model optimization technology.” Through these, we have succeeded in breaking through the limitations of conventional quantum computers.*4

These technologies will dramatically accelerate the industrial application of quantum computers in the “Early-FTQC era,” characterized by systems with tens of thousands of qubits.They hold the potential to make significant contributions to solving pressing societal challenges, such as shortening the development time for pharmaceuticals and new materials and improving the energy efficiency of chemical compound manufacturing. In this article, we will provide an easy-to-understand explanation of the details and future prospects of these innovative technologies. While it is a bit lengthy, if you read through to the end, the future of quantum computing will surely become much clearer!

A New Technology Enabling Earlier Practical Quantum Computing: STAR Architecture ver. 3
Dramatically Accelerating Molecular Energy Calculations for the Early-FTQC Era
- Background: Molecular Model Decomposition and Two Computational Approaches
- The Key Idea: Optimizing the Molecular Model Itself
Large-Scale Quantum Chemistry Simulation with STAR Architecture ver. 3 × Molecular Model Optimization
Summary and Future Outlook

A New Technology Enabling Earlier Practical Quantum Computing: STAR Architecture ver. 3

Below, we explain as plainly as possible the core technology behind our latest press release, STAR Architecture ver. 3. The key idea is to refine the basic set of primitive logical operations (called the universal gate set) used by fault-tolerant quantum computing (FTQC) architecture, so that it can switch to the most suitable operations for a given computation task. As a result, the STAR architecture's computational capabilities have improved dramatically, paving the way for quantum computers to be used for industrially relevant tasks much earlier.

What is a "universal gate set"?

Quantum computation can be expressed as a type of complex circuit comprising a small number of basic logical operations, known as quantum gates. These elementary operations are analogous to those in classical computing, such as adding or multiplying a single bit. Interestingly, it is well known that, in quantum computing, a specific set of quantum gates — the universal gate set — can be combined to perform complex computational tasks that are hard to execute on classical computers. A key finding of our study is that using an unconventional universal gate set can significantly impact the size of the executable circuit, its execution time and accuracy, and the number of qubits required.

In usual FTQC architectures, the following gate set is typically assumed (here, you do not need to know the meaning of each symbol):

$G_\rm{FTQC}=\lbrace \rm{CNOT},\it{H,S,T}\rbrace$

A particularly important gate in our studies is the T gate. It is a logical operation that rotates the phase of a qubit by a fixed angle, and mathematically, it can be written as the following (2×2)-matrix:

You can regard this as a “digital" rotation gate whose angle is fixed at π/4. By contrast, some representative quantum applications, such as materials simulation, require a large number of “analog” rotation gates at various finer angles, not just π/4. Mathematically, such analog operations can be written as

Here, the parameter $\theta$ can be varied continuously. In summary, the T gate is a digital rotation with a fixed angle, whereas $R(\theta)$ is an analog rotation whose angle can be chosen arbitrarily.

The conventional FTQC: Universally powerful, but expensive

In conventional FTQC, an arbitrary-angle rotation $R(\theta)$ is not executed directly. Instead, it is typically approximated by synthesizing many T gates together with other gates such as S and H, an approach known as gate synthesis. Schematically, it can be represented as

$R(\theta) \approx HTSHTSHTSHT…THTH$

This approach is highly versatile and theoretically very powerful. Its main drawback, however, is that it requires the execution of large numbers of high-fidelity T gates. This could be a serious bottleneck because the T gate is generally the most resource-intensive operation in FTQC. This is because, in most FTQC architectures, implementing a high-fidelity T gate relies on a costly preprocessing procedure known as magic state distillation.

In FTQC, a T gate is executed by consuming a special logical qubit known as a magic state. The most standard technique to prepare this state with high fidelity is magic state distillation. Roughly speaking, it is a method for producing a small number of high-quality magic states by gathering and consuming many noisy ones. More specifically, the magic state $|T\rangle$ corresponding to the T gate is described as follows:

$|T\rangle=T|+\rangle$

$| \rangle$ represents the quantum state, and $|+\rangle = \frac{|0\rangle + |1\rangle}{\sqrt{2}}$ represents the superpotition of $|0\rangle$ and $|1\rangle$ .

In practice, however, it is difficult to prepare a perfect magic state from the outset on a real quantum device. Therefore, one needs to begin with many noisy magic states and combine them to distill cleaner ones.

Magic state distillation is a remarkably powerful and general approach, but it also has a clear weakness. It requires a module akin to a dedicated “factory" for magic states - with a large number of qubits and separate from the processor itself. In other words, the magic state distillation is easy to adapt to many kinds of quantum tasks, but the required number of qubits becomes extremely large. Notably, this could be a serious challenge in the Early-FTQC era, when the number of available qubits is strictly limited.

The idea of STAR architecture: Realizing analog rotations at low cost

The Space-Time Efficient Analog Rotation quantum computing (STAR) architecture *5, developed jointly by Fujitsu and Osaka University in 2023*6, takes a fundamentally different approach. Rather than approximating analog rotations by synthesizing many T gates, it aims to execute the analog rotation gate $R(\theta)$ more directly, faster, and with fewer resources. As shown in Figure 1, the gate set handled by STAR architecture ver. 1 / ver. 2 is

$G_\rm{STAR}=\lbrace \rm{CNOT},\it{H,S,R(\theta)}\rbrace$

That is, the essential idea is to move away from the T-gate-centric approach that dominated conventional FTQC architecture and toward a gate set that puts arbitrary-angle analog rotations front and center. More specifically, there are several concrete ways to execute analog rotation gates directly, and the performance differences among them define the distinction between STAR ver. 1 and 2.

In particular, the method adopted in STAR architecture ver. 2 is well suited to quantum circuits that use very large numbers of small-angle rotations, and last year we demonstrated that it offers major advantages for some materials simulation tasks.

Figure 1: Differences among the logical gate sets used in conventional FTQC, STAR ver. 1/2, and STAR ver. 3. In ver. 3, the important point is that both analog rotation gates and logical T gates are available.

However, the STAR architecture also had a weakness

At the same time, earlier versions of the STAR architecture - particularly ver. 2 - had a clear weakness. As the rotation angle of the analog rotation gate grows larger, both execution speed and accuracy degrade significantly. Therefore, the STAR architecture is extremely powerful for quantum circuits dominated by small-angle rotation gates, but it has serious limitations for different types of quantum circuits.

Even when executing small-angle rotation gates, the issue above can still arise. In general, when the STAR architecture tries to execute R(θ) directly, it realizes the desired gate $R(\theta)$ with a probability of 1/2, but with the remaining probability 1/2 it instead produces the inverse rotation $R(-\theta)$ . When such a failure event occurs, one must double the rotation angle and execute R(2θ) again, using the relation $R(\theta)=R(2\theta)\cdot R(-\theta)$ , to correct the operation back to the target rotation. Of course, $R(2\theta)$ itself also fails with probability 1/2, so in that case the angle is doubled once more and fed back. By repeating this feedback process, the STAR architecture is guaranteed eventually to realize the target rotation angle (this is called a repeat-until-success, or RUS, process). Because the rotation angle grows exponentially during this process, an unlucky long RUS sequence forces feedback at large rotation angles. These large-angle rotations became a major source of logical errors in STAR architecture ver. 2, severely limiting the size of executable circuits.

Magic state cultivation rewrites the rules

Around the same time as the proposal of STAR architecture ver. 2, a new technique called magic state cultivation*7 emerged and attracted considerable attention in recent years. Similar to magic state distillation, this method is used to generate high-quality magic states required for T gates. However, conceptually, it is more akin to carefully growing the desired states step by step within a relatively small area than to mass production in a large factory. Broadly speaking, the difference between distillation and cultivation can be described as follows:

Distillation: gather many noisy magic states, entangle them, and distill only a few high-quality states within a large factory space
Cultivation: grow a clean single magic state gradually within a small space

Of course, this is only an intuitive explanation, but it is sufficient for readers of this blog. The important point is that cultivation makes it possible to prepare high-fidelity T gates far more compactly than before. This new technique enables the conventional FTQC approach to implement T gates with fewer qubits, while at the same time partially threatening the technical advantage previously held by the STAR architecture. Even so, the STAR architecture still retains a strength in implementing analog rotations efficiently, allowing small-angle rotation gates to be executed faster with fewer qubits and higher fidelity.

The core of STAR ver. 3: A paradigm shift in the universal gate set

The essence of our new quantum computing architecture (called STAR architecture ver. 3) lies here (for details, please refer to the original paper *8 ). In prior studies, there were two competing ideas:

The FTQC approach: centered on T gates, highly versatile, but requiring an enormous number of qubits for preparing T-gates.
The STAR approach: centered on analog rotation gates, fast and resource-efficient, but weak for some classes of quantum circuits.

In contrast, the STAR architecture ver. 3 has broken out of this either-or situation by successfully integrating magic state cultivation with the STAR-based approach. In other words, by adopting a new gate set

$G_\rm{STAR v3}=\lbrace \rm{CNOT},\it{H,S,R(\theta), T}\rbrace$

Consequently, the STAR architecture has evolved into a more general-purpose quantum computing machine that can exploit both analog rotation gates and logical T gates.

This corresponds to the right-hand side of Figure 1. Conventional FTQC uses T gates as the basic building block from which analog rotations are synthesized, while STAR ver. 1 / ver. 2 used $R(\theta)$ directly as the basic building block. By contrast, ver. 3 combines both "a direct way to perform analog rotations" and "a robust way to synthesize them with T gates." As discussed below, it is best understood not as adding one new tool to the quantum-computing toolbox, but as highly integrating two toolboxes designed for different purposes.

STAR-magic mutation: Getting the Best from Two Approaches

The unified gate set above enables faster and more accurate implementations of analog rotation gates through a new technique called STAR-magic mutation. The key idea is very straightforward: first try STAR-based rotation gate quickly, and if the RUS process looks likely to drag on, switch partway through to T-gate synthesis using cultivated magic states. Symbolically, the idea is described as follows:

$n \leq n_{th}$ $\Rightarrow$ STAR-based approach
$n > n_{th}$ $\Rightarrow$ Cultivation-based approach

Here, $n$ is the number of trials in the RUS process, and $n_th$ is the threshold for switching. This novel technique enables the STAR architecture ver. 3 to improve the logical error rate of analog rotation gates by roughly 1/10 to 1/1000 relative to the previous STAR architecture ver. 2, while simultaneously improving both the execution speed and the fidelity of analog rotations by roughly 100 times relative to conventional FTQC. In addition, by expanding the gate set, it became possible to execute a much broader class of quantum circuits. Figure 2 shows the scale of quantum circuits that can be executed within each quantum computing architecture.

Figure 2: Comparison of executable circuit size (number of gates) for conventional FTQC (Cultivation), STAR ver. 1/2, and STAR ver. 3. The vertical axis represents the number of executable analog rotation gates, and the horizontal axis represents the number of executable T gates.

Summary: STAR architecture ver. 3 is a redesign of the universal gate set

In a nutshell, the STAR architecture ver. 3 is "a new quantum computing architecture that lets analog rotation gates and T gates coexist and selects the most efficient execution method according to the target circuits." The conventional FTQC architecture was a powerful approach that assembled everything around the T gate. Earlier STAR architectures, by contrast, achieved speed and resource efficiency by executing analog rotations directly. STAR architecture ver. 3 does not place these two approaches in opposition; instead, it moves toward combining and selectively using them together.　In this sense, the advance realized by STAR architecture ver. 3 is not merely a performance improvement. Its value lies in taking the design of quantum computing architecture one step further. In conclusion, STAR architecture ver. 3 offers a promising route to solving industrially relevant problems with limited qubits, high accuracy, and practical speed.

Dramatically Accelerating Molecular Energy Calculations for the Early-FTQC Era

One of the most promising applications of quantum computers is the calculation of molecular energies. Strongly correlated molecular systems — which are critical to drug discovery and catalyst design — are particularly difficult to handle on classical computers, as exact simulations are often rendered infeasible by memory constraints. In this post, we introduce a set of novel optimization techniques that enable practical molecular energy calculations on early fault-tolerant quantum computers (Early-FTQC) using realistic computational resources.

Background: Molecular Model Decomposition and Two Computational Approaches

To compute molecular energies on a quantum computer, we first decompose the quantity representing the molecular energy — the Hamiltonian — into a large number of "combinations of simple operations." In mathematical terms, this takes the following form:

$\displaystyle \hat{H} = \sum_{l=1}^{L} c_l \hat{P_l}$

Here, each term has an associated weight $c_l$ representing its importance, and the basic operations executable on a quantum computer, $\hat{P}_l$ , are known as Pauli operators. By implementing the time evolution $e^{it\hat{H}}$ using this decomposition, we can leverage its properties to estimate molecular energies with high accuracy — a technique known as quantum phase estimation. Depending on the weight of each term $|c_l|$ , different implementation strategies can be applied:

Time evolution method (Trotterization) — Produces larger circuits, but execute deterministically.
Random sampling method — A lightweight approach that probabilistically selects terms to execute. Circuits are smaller, but repeated execution is required.

For molecular Hamiltonians, it is known that applying the first method to high-weight terms and the second to low-weight terms reduces the overall computational cost *9. However, conventional approaches treat the Hamiltonian as fixed, meaning optimization is limited to working within the given weight distribution.

The Key Idea: Optimizing the Molecular Model Itself

In this work, we propose a technique called Unitary Weight Concentration (UWC) that transforms only the representation of the Hamiltonian — without changing the underlying molecular energy — in order to optimize the weight distribution of its terms. (For full details, please refer to the original paper*10.) To make this concrete, let's use an analogy: imagine transporting cargo by car. Computing molecular energies corresponds to the mission of moving the cargo. In this analogy, the molecular model optimization (UWC) is like planning how to pack and transport the load. Even over the same distance, poor packing forces you to make multiple trips. But if you sort the heavy and light items efficiently, you can get everything done in fewer runs. UWC corresponds to this "cargo sorting" step: it reduces unnecessary trips without changing what needs to be transported (i.e., the molecular energy structure). More concretely, UWC transforms the Hamiltonian representation to restructure its coefficient distribution $|c_l|$ , thereby:

Rebalancing the ratio of large and small terms, and
Optimizing the division of labor between the time evolution method and the random sampling method.

It is worth noting that this optimization is performed on classical computers or high-performance computing (HPC) systems. In other words, this approach maximizes overall performance by combining classical and quantum computation. This reduces the total gate count in the quantum circuit — enabling resource savings that go beyond ordinary circuit optimization by intervening at the level of the algorithm's structure itself.

Figure 3: Conceptual diagram of molecular model optimization (UWC)

Large-Scale Quantum Chemistry Simulation with STAR Architecture ver. 3 × Molecular Model Optimization

Target Molecules

To validate the effectiveness of this technology, we performed resource estimations for the following industrially important molecules. (For full details, please refer to the original paper.) All three systems are strongly correlated and of a scale that makes exact classical simulation infeasible. They also represent problem sizes that were beyond the reach of the previous STAR architecture ver. 2.

Cytochrome P450 — A family of enzymes in the liver responsible for metabolizing drugs and toxins, and critical for understanding side effects and variability in drug response in pharmaceutical research. For this study, we used a 62-spin-orbital model of the most reactive intermediate *11.
Iron–sulfur cluster — The active center of proteins involved in many biological processes, including photosynthesis, respiration, DNA repair, and metabolism. It has also attracted attention for its connection to green ammonia synthesis. For this study, we used a 72-spin-orbital model containing four iron and sulfur atoms *12.
Ruthenium catalyst — A catalyst used across a wide range of applications including organic synthesis, petrochemicals, industrial processes, and fuel cells. For this study, we used a 58-spin-orbital model of a ruthenium catalyst that reduces CO₂ to ethanol fuel — a process relevant to decarbonization *13.

Figure 4: Structural models of the three molecules examined in this study

Qubit Reduction and Improved Logical Error Tolerance with STAR Architecture ver. 3

Our resource estimation results show that the number of physical qubits required can be reduced by approximately one to two orders of magnitude compared to conventional fault-tolerant quantum computing (FTQC). Traditional FTQC approaches require a mechanism known as "magic state factories," which consume a large number of qubits, along with many additional ancilla qubits. The STAR architecture significantly reduces the need for these extra resources, leading to a substantial overall reduction in qubit count. Furthermore, while the STAR architecture ver. 2 required a physical error rate of 0.01%, the STAR architecture ver. 3 is capable of operating at 0.1%. This reflects a major improvement in logical error tolerance. Notably, the 0.1% threshold has already been achieved experimentally, suggesting that practical quantum computation in the Early-FTQC era is becoming an increasingly realistic prospect.

Figure 5: Number of physical qubits required for energy calculations of the three target molecules

Dramatic Reduction in Computation Time via Molecular Model Optimization (UWC)

The molecular model optimization (UWC) technique delivers a further dramatic speedup: compared to running without it, the estimated computation time for molecular energy calculations on STAR architecture ver. 3 is reduced by approximately 1/1000. Concretely, calculations previously estimated to take several decades or more have been brought down to a matter of weeks through the gate count reduction enabled by UWC. These results assume a physical error rate of 0.1%, which is considered realistic for current devices.

Table: Estimated execution times for each molecule at physical error rates of 0.1% and 0.01%. Values in parentheses correspond to the 0.01% case.

	Cytochrome P450	Iron–sulfur cluster	Ruthenium catalyst
STAR ver. 3	36 years（14 years）	105 years（34 years）	39 years（14 years）
STAR ver. 3×UWC	36 days（8.9 days）	42 days（13 days）	28 days（4.6 days）

Looking ahead, further reductions in execution time can be expected as quantum hardware continues to improve. Parallel execution across multiple quantum computers is also possible. Taken together, these factors suggest that computation on a practical timescale is becominga realistic goal.

Figure 6: Estimated computation time for energy calculations of the three target molecules

Significance: A Practical Roadmap for Quantum Computing in Drug Discovery and Materials Science

The key contribution of this work is demonstrating that industrially meaningful problems can potentially be tackled using Early-FTQC hardware that is experimentally within reach — rather than requiring the large-scale quantum computers of the distant future. These results offer a concrete roadmap toward practical quantum computing applications in drug discovery, catalyst design, CO₂ reduction, and related fields.

Summary and Future Outlook

In this article, we have explained two key technologies: the STAR Architecture ver. 3 and molecular model optimization. The development of these technologies represents a major step forward in the practical application of quantum computers during the Early-FTQC era. By significantly reducing computational resources and enabling accurate energy calculations for chemical materials within a realistic timeframe, these technologies are expected to find applications across a wide range of fields, including drug discovery, new material development, and carbon recycling technologies. We will continue to further develop the STAR architecture and molecular model optimization technology, and accelerate our research and development to realize a future where quantum computers contribute to solving societal challenges. Please stay tuned for future developments in quantum computing! For more details, please refer to the press release below and the papers*8, *10 published on the same day.

Fujitsu and The University of Osaka develop new technologies for chemical material energy calculations on early-FTQC quantum computers

*1:Fujitsu Quantum

*2:Center for Quantum Information and Quantum Biology (QIQB), The University of Osaka

*3:Fujitsu Small Research Lab

*4:Fujitsu and The University of Osaka develop new technologies for chemical material energy calculations on early-FTQC quantum computers [Press Release: March 25, 2026]

*5:Dedicated page for STAR architecture

*6:Fujitsu and Osaka University develop new quantum computing architecture, accelerating progress toward practical application of quantum computers [Press Release: March 23, 2023]

*7:magic state cultivation

*8:R. Toshio et al., "STAR-Magic Mutation: Even More Efficient Analog Rotation Gates for Early Fault-Tolerant Quantum Computer," arXiv:2603.22891 (2026).

*9:J. Günther et al., "Phase estimation with partially randomized time evolution", arXiv:2503:05647

*10:S. Kanasugi et al., "Enabling Chemically Accurate Quantum Phase Estimation in the Early Fault-Tolerant Regime," arXiv:2603.22778 (2026)

*11:J. J. Goings et al. "Reliably assessing the electronic structure of cytochrome P450 on today’s classical computers and tomorrow’s quantum computers", Proc. Natl. Acad. Sci. U. S. A. 119, e2203533119 (2022).