Giuseppe Carleo
@gppcarleo.bsky.social
Computational Quantum Physicist - EPFL Lausanne, Switzerland
congratulazioni!
September 24, 2025 at 1:03 PM
congratulazioni!
The training framework is certainly important (for example one needs to use the "correct version" of SR in this case, but also the neural networks should be good enough to generalize. For Ising it definitely works with transformers ! arxiv.org/abs/2502.09488
Foundation Neural-Network Quantum States
Foundation models are highly versatile neural-network architectures capable of processing different data types, such as text and images, and generalizing across various tasks like classification and g...
arxiv.org
April 12, 2025 at 7:58 PM
The training framework is certainly important (for example one needs to use the "correct version" of SR in this case, but also the neural networks should be good enough to generalize. For Ising it definitely works with transformers ! arxiv.org/abs/2502.09488
A foundation NQS allows to reduce (by several orders of magnitude!!) the cost needed in QMC/Diffusion Monte Carlo calculations to span the same phase diagram (while achieving higher accuracy than DMC). We also include, at T=0, the full quantum wave functions for protons, beyond Born-Opp. (2/2)
Universal neural wave functions for high-pressure hydrogen
We leverage the power of neural quantum states to describe the ground state wave function of solid and liquid dense hydrogen, including both electronic and protonic degrees of freedom. For static prot...
arxiv.org
April 11, 2025 at 9:05 AM
A foundation NQS allows to reduce (by several orders of magnitude!!) the cost needed in QMC/Diffusion Monte Carlo calculations to span the same phase diagram (while achieving higher accuracy than DMC). We also include, at T=0, the full quantum wave functions for protons, beyond Born-Opp. (2/2)
Read the full preprint here: arxiv.org/abs/2503.14502 5/5
Precise Quantum Chemistry calculations with few Slater Determinants
Slater determinants have underpinned quantum chemistry for nearly a century, yet their full potential has remained challenging to exploit. In this work, we show that a variational wavefunction compose...
arxiv.org
March 20, 2025 at 8:45 AM
Read the full preprint here: arxiv.org/abs/2503.14502 5/5
Our approach surpasses all second-quantized NQS results for molecules published so far, despite being a much simpler ansatz conceptually. This suggests that for small to intermediate molecules, fully correlated wave functions might not be necessary. 4/5
March 20, 2025 at 8:45 AM
Our approach surpasses all second-quantized NQS results for molecules published so far, despite being a much simpler ansatz conceptually. This suggests that for small to intermediate molecules, fully correlated wave functions might not be necessary. 4/5
Using optimized contractions, our method scales computational cost with the fourth power of the number of basis functions. Benchmarking against exact full-configuration interaction results, we achieved lower variational energies than CCSD(T) for several molecules in the double-zeta basis. 3/5
March 20, 2025 at 8:45 AM
Using optimized contractions, our method scales computational cost with the fourth power of the number of basis functions. Benchmarking against exact full-configuration interaction results, we achieved lower variational energies than CCSD(T) for several molecules in the double-zeta basis. 3/5
While this ansatz is as old as quantum chemistry, fully optimizing it has been challenging. Our innovation lies in efficiently optimizing the determinants by leveraging the quadratic dependence of energy on selected parameters, allowing for exact optimization. 2/5
March 20, 2025 at 8:45 AM
While this ansatz is as old as quantum chemistry, fully optimizing it has been challenging. Our innovation lies in efficiently optimizing the determinants by leveraging the quadratic dependence of energy on selected parameters, allowing for exact optimization. 2/5
In the absence of a community consensus on what it really means to obtain quantum advantage over **all possible** classical methods, why keep stirring controversy instead of stating that advantage is over some X or Y classical numerical methods "only"?(2/2)
March 13, 2025 at 7:24 PM
In the absence of a community consensus on what it really means to obtain quantum advantage over **all possible** classical methods, why keep stirring controversy instead of stating that advantage is over some X or Y classical numerical methods "only"?(2/2)
I am insinuating nothing. the preprint is pretty clear we analyze the diamond geometry. we will provide more geometries soon, but the burden of proof that you can beat "all classical methods" is on your side, since you decided to make this questionable (cynical?!) claim in the first place.
March 13, 2025 at 7:07 PM
I am insinuating nothing. the preprint is pretty clear we analyze the diamond geometry. we will provide more geometries soon, but the burden of proof that you can beat "all classical methods" is on your side, since you decided to make this questionable (cynical?!) claim in the first place.
we compared against ground truth computed with MPS at the largest scale you provided, and also against your own experiment at the scale where mps was not available... I'm not sure what your remark is about
March 13, 2025 at 6:45 PM
we compared against ground truth computed with MPS at the largest scale you provided, and also against your own experiment at the scale where mps was not available... I'm not sure what your remark is about
If they are easy as you suggest, why you claimed they would take ~200 years on one of the largest supercomputers available ? I'm confused.
March 13, 2025 at 6:43 PM
If they are easy as you suggest, why you claimed they would take ~200 years on one of the largest supercomputers available ? I'm confused.