An improved method of estimating molecular weights of volatile organic compounds from their mass spectra has been developed and evaluated for accuracy. This technique can be implemented with a user friendly expert system on a personal computer. The method is based on a pattern recognition based classifier and empirically derived rules from a set of toxic and related organic compounds. The set of compounds consists of six classes: nonhalobenzenes; chlorobenzenes; bromo- and bromochloroalkanes/alkenes; mono- and dichloroalkanes/alkenes; tri, tetra- and pentachloroalkanes/alkenes; and unknowns. In the present study improvements for the unknown class, which is the largest class in actual samples, were made. For this latter class new rules were derived using a new feature of the mass spectrum. Extensive testing of the system was conducted with reference and field data. The probable errors in the estimated molecular weights were 0.5 daltons for training reference spectra and 5.0 and 1.3 daltons for randomly selected and field spectra, respectively. These errors are about one-half those of previous versions of the system. The test results show that the domain of validity of this present system greatly exceeds the original training domain.