Down memory lane: Quantum Computing

I spent about 2.5 years working on variational quantum algorithms for noisy intermediate-scale quantum (NISQ) devices [1]. The question was straightforward: can we do anything useful with these noisy, small-scale, "universal" quantum devices? Here useful was typically to mean faster, but could also mean solve far too complex problems for classical quantum chemistry.

The short answer I came to was: not in any way that clearly demonstrated a speedup over well-tuned classical solvers. The longer answer: you can get them to run, get numbers back, even match the literature, but there's no clear, reproducible speed-up. Most of the effort is in engineering the system to produce anything coherent, not in pushing computational frontiers. I'm not sure if this is a good thing or a bad thing.

So what is the state now? Are there any clear signs of utility for variational quantum algorithms? Are there any new quantum algorithms for chemistry or physics that require error correction but have proven advantage over classical computers? My guess is the answer is no not really, but I'm not sure and for now don't have the time to read up and investigate.

What I saw with Universal Quantum Computing in the NISQ Era

Universal quantum computing here means we have a qubit, quantum information analog of digital bits, and we can do arbitrary single-qubit rotations and controlled two-qubit gates to produce logic operations that put qubits into superposition and can entangle them. We can also compose them into circuits that approximate any unitary evolution.

For famous algorithms like Shor's or Grover's, the path to usefulness is clear if you had fault-tolerant quantum computing with sufficient qubits. But in the NISQ setting, VQA [2] or QAOA are the only viable options. There might be some new class of NISQ-like algorithms, I'm not sure, but it's probably still safe to say VQA and QAOA are dominant.

Lets say that you moved beyond the NISQ era, which may well be happening, and thus the noise and decoherence are under control, there's still the ansatz problem, that is how do you pick a circuit structure that can efficiently represent the solution state in the first place?

The Ansatz Problem

An ansatz is a parameterized (quantum circuit) guess for the form of your quantum state. In VQAs, it's the fixed sequence of gates you tune with a classical optimizer. In phase estimation (PEA) or Hamiltonian simulation, it's often the state-preparation step for your quantum algorithm.

The difficulty is balancing expressivity and feasibility:

• Too shallow: can't represent the physics; optimizer converges to the wrong state.
• Too deep: hardware noise kills you in NISQ; in fault-tolerant hardware, depth inflates T-gate and qubit costs.
• Too generic: risks barren plateaus [3].
• Too problem-specific: works only on one Hamiltonian.

In PEA, the ansatz problem just shifts to the state-preparation step. You might nail the controlled-unitary and inverse QFT, but if you can't efficiently prepare the eigenstate, you'll likely get garbage.

So what did I work on?

Most of the creativity in the research I did comes all from my colleague/co-author, I was mostly involved in domain application, instrumentation, and analysis. There were two works [4-5] but the one I'll highlight is probably the least interesting: we developed a hybrid quantum–classical eigensolver without variation or parametric gates[4]. The idea is to project the problem Hamiltonian into a smaller subspace, measured term-by-term with short circuits, and diagonalized classically.

This allowed us to:

• Extract ground and excited states for small molecules (BeH₂, LiH).
• Validate against exact diagonalization.
• Run on the quantum hardware at the time (i.e., IBM devices).

It avoided long, problem-tailored ansatz circuits, but the choice of basis in the subspace projection is still a hidden ansatz.

A Self-Critique of Our Hybrid Eigensolver Work

We avoided variational loops and deep, problem-specific ansätze: no parameterized circuits, no barren plateau optimizers. The issue, though, was

1. We dodged the ansatz problem: The reduced-space basis is still an ansatz, thus performance depends on making a smart choice. We didn't quantify sensitivity.
2. Hardware vs. simulation gap unexplored IBM runs matched noiseless simulations, but was this due to noise resilience, shallow circuits, or luck?
3. Thin classical comparisons We used exact diagonalization due to simplicty of the chemical system and basis. Real claims require benchmarks vs. DMRG [6], coupled-cluster, etc.

Why NISQ VQAs Struggle

There has been a lot of work on this in the last few years and I'm not fully up to date but this is what I gather:

• Noise vs. depth: deeper means more decoherence.
• Barren plateaus: gradients vanish exponentially with qubits.
• Optimizer instability: hardware drift, shot noise, optimizer quirks.
• Classical competition: tensor networks, DMRG [6] often scale better [7].
• Ansatz rigidity: wrong ansatz wastes all gates and shots.

Summary of My Position

I'm not a seasoned quantum algorithm researcher, but from my limited research, I see NISQ-era of VQAs as mainly useful for benchmarking, with their progress limited by the challenge of designing effective ansätze. In quantum chemistry, there is little convincing evidence so far that VQAs offer practical advantages¹. Digital quantum simulation is highly flexible but comes with significant resource costs. Analog simulation, which directly emulates physical systems, is already useful in certain specialized areas but will always be a niche. Looking ahead to fault-tolerant quantum computing, real breakthroughs may be possible, but efficient state preparation will remain a central obstacle.

Footnotes

---

1. Maybe this has changed but I wager probably not and the paper by Lee et al. [7] is a good indicator. ↩

References

[1] J. Preskill, Quantum Computing in the NISQ era and beyond, arXiv (2018). [2] M. Cerezo et al., Variational Quantum Algorithms, arXiv (2021). [3] M. Larocca et al., Barren Plateaus in Variational Quantum Computing, arXiv (2024). [4] P. Jouzdani & S. Bringuier, Hybrid Quantum–Classical Eigensolver Without Variation or Parametric Gates, Quantum Reports 3, 8 (2021). DOI [5] P. Jouzdani, S. Bringuier, M. Kostuk, A method of determining molecular excited-states using quantum computation, MRS Advances 6 (2021) 558–563. DOI [6] S. R. White, Density matrix formulation for quantum renormalization groups, PRL 69, 2863 (1992). [7] S. Lee et al., Evaluating the evidence for exponential quantum advantage in ground-state quantum chemistry, Nat. Commun. 14, 1952 (2023). DOI

Reuse and Attribution

A Year for Robotics in Labs?

Since 2022 the focus of AI has been mostly driven by large language models(LLMS) (LLMs) and generative AI and advancing them to concepts like artificial general intelligence (AGI) and eventually artificial superintelligence (ASI). I don't know what the time frame looks like for AGI and ASI or even if they are achievable with transformer and reinforcement learning techniques, given its not clear if these work because of data scaling or reward policy. What I'm more interested in for this year is I think the maturity of LLMs and the onset of agentic AI we are going to see a lot more activity in autonomous robotics. As a consequence, the availability of autonomous robotics tied to a materials synthesis and characterization lab will be come more more practical.

This shift from human in the lab to robot in the lab is going to change how scientist conduct their work and experiments. The reason being is small R&D iterative experiments will rely less on human intuition and more on data driven approaches that guide autonomous robots and systems.

Checkerboard of autonomy levels for labs (adapted from figure 1 in ref. [5])

Due to the nature of chemistry synthesis and characterization, some of the most early examples of autonomous robotics occurred with optimizing analytical procedures for spectroscopic detection [5]. These systems performed low supervision (i.e., limited human intervention) experimental tasks planned by data-driven approaches, such as maximizing the intensity of a characteristic absorbance peak. However, we are now entering an era, at least I believe, where self-driving labs (SDLs) will become a staple in small R&D labs.

Series Focus 🤖🥼

This is a series of posts on the topic of self-driving labs that I write about this year, at least thats the plan. The particular focus of this post is more on the robotic manipulator aspect of SDLs as opposed to automation of equipment through microcontrollers or software. For example automating the control of an SEM through IO interfaces would not not be a robotics implementation, unless your utilizing a humanoid robot to interface with the keyboard/mouse to perform the task; not something I'm really referencing in this post nor do I think we are anywhere close to that.

Objectives for Robotics in Labs

In short the basic idea of autonomous robotics in the lab and thereby self-driving laboratories (SDLs)¹ revolves around the optimization of robotic manipulators' actions through a data-driven approach. The systems form a closed loop where data is continuously collected, analyzed, and used to guide the next set of actions. For instance, in the work by Abolhasani et al. [1], AI/ML are integrated into SDLs to form hypotheses for experimentation, with the output being shared datasets and publications. This closed-loop process ensures that data analysis informs the actions of robotic manipulators in the SDL, optimizing the experimental workflow.

Another approach, as discussed by Bai et al. [2], is a goal-driven architecture where researchers set goals/objectives and resource restrictions², triggering a closed-loop process. The use of AI/ML propose new experiments, translating conditions into machine-actionable recipes that control the hardware for reaction and characterization. This iterative process ensures that the system's actions are continuously optimized based on data analysis and comparison to research goals.

To give some insight to an inorganic materials problem, Chen et al. [3] demonstrated a thermodynamic strategy in a robotic enabled synthesis lab. The SDL navigates high-dimensional phase diagrams to maximize reaction energy and achieve desired phase purity. The AutoSciLab framework by Desai et al. [4], uses active learning³ to efficiently design experiments by creating hypotheses in a latent space. The use of active learning enables the system to "focus" on the most informative data points in the experimental space. As a result this can greatly reduce the number of experiments needed to achieve a desired result.

These SDLs aren't without their challenges and the robotic manipulation although amazing still will need to have the close to the human capabilities of being nimble and adjustable. Current robotic manipulators have significant degrees-of-freedom and fine motor capabilities. However, the ability to perform complex dynamic tasks with high precision and adaptability is still challenging. There is also the need to create benchmark and performance metrics for SDLs as highlighted by Volk et al. [6]. The authors emphasized the importance of reporting parameters and performance metrics to guide research in SDLs. This is absolutely necessary to be able to compare the performance of different SDLs. This data is needed to aid and inform decision-making of future SDL implementations.

While the terminology may vary for autonomous robotics in labs, the overarching objective remains consistent: to delineate an autonomous robotic process that leverages data-driven methodologies to optimize specific experimental goals. Examples of this include maximizing reaction energy, achieving phase purity, and enhancing spectroscopic signals. By employing closed-loop systems, SDLs continuously collect and analyze data, allowing robotic manipulators to adapt their actions based on real-time insights. This iterative approach not only streamlines the experimental workflow but also fosters a more efficient and effective research environment, ultimately advancing the landscape of materials synthesis and characterization.

Footnotes

---

1. A pretty good list of review and original research papers on self-driving labs can be found at awesome-self-driving-labs. Its not comprehensive but its a good start. ↩

2. This is important because despite a SDL or autonomous systems potential for carrying out actions, it is very much the case that the actual repeated resource(s) is limited. Here resource could be a very expensive precursor for synthesis or timely characterization. Another way to think of this is as the time the robot has to perform the experiment, the amount of energy the robot has to perform the experiment, the amount of money the robot has to perform the experiment, etc. ↩

3. Active learning is a ML technique where a model is trained by selecting (through external query) the most informative data points for learning, where as traditional ML approaches that use a fixed dataset. ↩

References

[1] M. Abolhasani, K.A. Brown, Guest Editors, Role of AI in experimental materials science, MRS Bulletin 48 (2023) 134–141. DOI.

[2] J. Bai, S. Mosbach, C.J. Taylor, D. Karan, K.F. Lee, S.D. Rihm, J. Akroyd, A.A. Lapkin, M. Kraft, A dynamic knowledge graph approach to distributed self-driving laboratories, Nat Commun 15 (2024) 462. DOI.

[3] J. Chen, S.R. Cross, L.J. Miara, J.-J. Cho, Y. Wang, W. Sun, Navigating phase diagram complexity to guide robotic inorganic materials synthesis, Nat. Synth 3 (2024) 606–614. DOI.

[4] S. Desai, S. Addamane, J.Y. Tsao, I. Brener, L.P. Swiler, R. Dingreville, P.P. Iyer, AutoSciLab: A Self-Driving Laboratory For Interpretable Scientific Discovery, (2024). DOI.

[5] G. Tom, S.P. Schmid, S.G. Baird, Y. Cao, K. Darvish, H. Hao, S. Lo, S. Pablo-García, E.M. Rajaonson, M. Skreta, N. Yoshikawa, S. Corapi, G.D. Akkoc, F. Strieth-Kalthoff, M. Seifrid, A. Aspuru-Guzik, Self-Driving Laboratories for Chemistry and Materials Science, Chem. Rev. 124 (2024) 9633–9732. DOI.

[6] A.A. Volk, M. Abolhasani, Performance metrics to unleash the power of self-driving labs in chemistry and materials science, Nat Commun 15 (2024) 1378. DOI.

---

Reuse and Attribution