Vous êtes ici : Accueil > Laboratoire PHELIQS > Techniques de réseaux tensoriels pour la simulation de circuits supraconducteurs quantiques

agenda


Soutenance de thèse

Techniques de réseaux tensoriels pour la simulation de circuits supraconducteurs quantiques

​​​​
​​​​​​Adrien Moulinas
​​
14/12/2026 @ 9​​h

CEA - Grenoble
Bâtiment 10.05
Salle 445
​
Publié le 14 décembre 2026

Adrien Moulinas

Abstract: The goal of superconducting quantum hardware design is to build devices whose quantum operations reach the highest possible fidelity. This fidelity is limited by effects such as leakage to higher energy levels, dissipation and decoherence, and designing better devices therefore requires simulation tools that can account for these deviations from the ideal scenario. Unfortunately, the curse of dimensionality drastically limits our ability to simulate the Lindblad equation beyond the smallest systems: the memory needed to store the density matrix grows exponentially with the number of modes and with the size of each mode. In this thesis, we develop a tensor network based technique to simulate the Lindblad dynamics of a few interacting bosonic modes, with a focus on superconducting quantum circuits. ​​​​
​​​​
The technique uses three nested levels of compression. It detects and takes advantage of two very common situations: (i) the density matrix being pure or not far from pure and (ii) the entanglement between different modes being moderate (typically qubit-like). However, (iii) the occupation of the modes can be arbitrarily high, which makes naive truncations of Fock spaces (number of photons) inefficient. To exploit these features, (i) we work with a global purification of the density matrix, (ii) we compress the connection between different modes to account for the moderate entanglement and (iii) we use a quantics representation of the Fock occupation number, in which the occupation is written in binary and each bit is treated as a separate variable. Combining the three levels into a single tensor network framework brings drastic compression and speed-up of the simulations, it is the main contribution of this thesis. We build the quantics representation of the states and operators of a single mode, and we integrate the Lindblad equation in this format with schemes based on Kraus maps. ​​​​
​​​​
We illustrate the technique on two problems. For a Z gate on a cat qubit, the compressed state does not grow with the size of the cat, and we simulate cats of more than a thousand photons, whereas dense solvers, even on a GPU, stop at a few tens of photons. The state then takes a few tens of kilobytes of memory, against more than a terabyte for the dense density matrix. We also study how different error terms affect the gate error and the cost of the simulation. The ionisation of a transmon during readout is a more demanding regime, since the state becomes strongly mixed, the transmon and the cavity get entangled and both modes climb to high excitations, that we use to test how the compression behaves when its assumptions break down. In the regime for which it is designed, the memory and the cost of a time step grow only with the logarithm of the size of each mode. This favourable scaling should bring genuine computer-assisted design of these systems within scope.



​