notesum.ai

Published at October 30

FlowLLM: Flow Matching for Material Generation with Large Language Models as Base Distributions

cs.LG

cond-mat.mtrl-sci

cs.AI

stat.ML

Released Date: October 30, 2024

Authors: Anuroop Sriram¹, Benjamin Kurt Miller², Ricky T. Q. Chen¹, Brandon M. Wood¹

Aff.: ¹FAIR, Meta; ²University of Amsterdam

Arxiv: http://arxiv.org/abs/2410.23405v1

Refer to caption

Method	LLM Params	Integ. Steps	Validity (%) $\uparrow$		Coverage (%) $\uparrow$		Property $\downarrow$		Stability Rate (%) $\uparrow$	SUN Rate(%) $\uparrow$
			Structural	Composition	Recall	Precision	wdist ( $\rho$ )	wdist ( $N_{el}$ )	MP-2023
CDVAE [xie2021crystal]	–	5000	100.00	86.70	99.15	99.49	0.688	0.278	1.57	–
DiffCSP [jiao2023crystal]	–	1000	100.00	83.25	99.71	99.76	0.350	0.125	5.06	3.34
FlowMM [miller2024flowmm]	–	1000	96.85	83.19	99.49	99.58	0.239	0.083	4.65	2.34
CrystalLLM (70B) [gruver2024fine]	$\tau=0.7$	–	99.6	95.4	85.8	98.9	0.81	0.44	5.28	–
FlowLLM-Types	$\tau=0.5,P=0.9$	750	99.96	93.32	96.85	99.78	0.846	0.209	8.79	–
FlowLLM-Types	$\tau=0.9,P=0.9$	750	99.88	91.69	97.18	99.76	1.14	0.20	8.95	–
FlowLLM	$\tau=1.0,P=0.9$	250	99.81	89.05	99.06	99.68	0.66	0.09	10.07	4.89
FlowLLM	$\tau=0.7,P=1.0$	250	99.88	89.45	99.06	99.71	0.73	0.14	13.03	4.88
	$\tau=0.7,P=0.9$	250	99.94	90.84	96.95	99.82	1.14	0.15	17.82	4.92