Serious games have become increasingly available to educators. Empirical studies and meta-analyses have examined their impact on learning achievement. However, natural sciences could have a special relation to serious games by their systematic use of quantitative and predictive models that can generate microworlds and simulations. Since no known meta-analysis on serious games observed a significant impact in the specific context of science learning, the present meta-analysis synthesised results from 79 empirical studies that compared the impact on science learning achievement of instruction using serious games versus instruction using more conventional methods. Consistent with theory and past meta-analyses not specifically related to science learning, post-instruction learning achievement was weakly to moderately higher for declarative knowledge, knowledge retention and procedural knowledge for students taught with serious games. Furthermore, findings of the present work suggest that five moderator variables produced significant effects on the relationship between playing serious games and learning outcomes, and three showed consistent variations in mean effect size that could lead to significance, with more studies and larger samples. These findings are discussed in connection with previous meta-analyses' findings, potential pedagogical implications and possible future research.