Altered precipitation regimes and increased temperatures can enhance wildfire risk, particularly in semi-arid and arid ecosystems. Since these ecosystems occupy a third of the earth's land area, the immediate loss of carbon and the post fire reduction in carbon sequestration will significantly affect regional and global carbon budgets. However, there are few studies that quantify the recovery of ecosystem carbon production post fire in semi-arid ecosystems. In January 2014, a wildfire caused severe loss and damage to plants in a semi-arid Eucalyptus woodland in South Australia. The carbon flux of the woodland was monitored pre- and post-fire from an in-situ eddy covariance flux tower. Daily net ecosystem production (NEP), ecosystem respiration (Reco) and gross primary production (GPP), together with environmental variables collected in 2012 and 2013 (pre-fire) were used to generate machine learning regression algorithms. Subsequently, the algorithms were used to simulate Reco, GPP and NEP at the same site without fire from the environmental variables, which is referred to as modelled carbon fluxes. Changes in the difference between the modelled fluxes and the actual post-fire measured fluxes were used to investigate the recovery of Reco, GPP and NEP during four years after the fire (April 2014–March 2018). We found that the woodland was a carbon source throughout the first year and in the hot and dry months of the second year, but was a carbon sink in the third and fourth year after the fire. The recovery of NEP to pre-fire levels was estimated to take approximately 35 months (30–43 months at 95% confidence interval). Post fire measured GPP responded more strongly to rainfall events than indicated by the modelled values which suggests that carbon sequestration by new leaf growth is more responsive to rainfall than equivalent mature leaf area of this woodland. After four years, both measured Reco and GPP rates were still lower than modelled rates indicating that the mass of carbon and its turnover in the ecosystem was not yet at pre-fire levels. We conclude that full recovery from fire in this woodland, and likely in similar ecosystems, is strongly influenced by the incident rainfall regime after fire and will take five years or more, especially if total annual rainfall is less than the long term average.