Building on past studies showing that mobile learning improves learning outcomes and differs within a domain (e.g., science), this meta-analysis models domain-specific differences (e.g., learning activities) that drive these differences in science performance. A systematic database search (i.e., Web of Science, JSTOR, ERIC, PsycINFO, ProQuest Dissertations, and ACM) identified controlled experiments. A meta-analysis determined the overall effect of mobile learning on 4,145 primary and secondary school students' science achievement, and tested for moderator effects across 57 effect sizes from 44 studies. Mobile learning increased science achievement (g = 0.857). Mobile learning's effect sizes were larger for (a) inquiry or game-based learning, than other learning activities; (b) biology, and progressively smaller for earth and space sciences, chemistry, and physics; (c) activities jointly led by students and teachers, followed by those led by students; and then those led by teachers; (d) collectivistic countries than individualistic ones; and (e) primary or middle school students. Intervention duration, device type, learning environment, and publication year showed no moderation effects. Hence, a comprehensive theory of mobile learning must include age, learning approach, subject area, user roles, and cultural values. These results also suggest that science educators using mobile learning might improve student learning by (a) integrating it with science inquiry or game-based learning, (b) starting with biology before other science topics, (c) using learning activities jointly led by students and teachers, (d) starting with primary or middle school, and (e) starting with students in countries with collectivist cultural values to help one another.